JP7728899B2 - Synchronously replicating datasets and other managed objects to a cloud-based storage system - Google Patents

Description

[0001] FIG. 1 illustrates a first example system for data storage according to some embodiments. [0002] FIG. 1 illustrates a second example system for data storage according to some embodiments. [0003] FIG. 2 illustrates a third example system for data storage according to some embodiments. [0004] FIG. 2 illustrates a fourth example system for data storage according to some embodiments. [0005] FIG. 1 is a perspective view of a storage cluster having multiple storage nodes and internal storage coupled to each storage node to provide network-attached storage, according to some embodiments. [0006] FIG. 1 is a block diagram illustrating an interconnect switch coupling multiple storage nodes according to some embodiments. [0007] FIG. 1 is a multi-level block diagram illustrating the contents of a storage node and the contents of one of the non-volatile solid-state storage units according to some embodiments. [0008] FIG. 2 illustrates a storage server environment using some of the storage node and storage unit embodiments of the above figures, according to some embodiments. [0009] FIG. 1 is a blade hardware block diagram showing the control plane, compute plane, and storage plane, as well as authorities interacting with underlying physical resources, according to some embodiments. [0010] FIG. 2 illustrates an elasticity software layer on a blade of a storage cluster according to some embodiments. [0011] FIG. 2 illustrates an authority and storage resources at blades of a storage cluster, according to some embodiments. [0012] FIG. 1 illustrates a diagram of a storage system coupled for data communication with a cloud service provider in accordance with some embodiments of the present disclosure. [0013] FIG. 1 illustrates a diagram of a storage system according to some embodiments of the present disclosure. [0014] FIG. 1 illustrates a block diagram showing multiple storage systems supporting pods in accordance with some embodiments of the present disclosure. [0015] FIG. 1 illustrates a block diagram showing multiple storage systems supporting pods in accordance with some embodiments of the present disclosure. [0016] FIG. 1 illustrates a block diagram showing multiple storage systems supporting pods in accordance with some embodiments of the present disclosure. [0017] FIG. 2 illustrates a flowchart illustrating an example method for establishing a synchronous replication relationship between two or more storage systems in accordance with some embodiments of the present disclosure. [0018] FIG. 10 illustrates a flowchart depicting an additional example method for establishing a synchronous replication relationship between two or more storage systems in accordance with some embodiments of the present disclosure. [0019] FIG. 10 illustrates a flowchart depicting an additional example method for establishing a synchronous replication relationship between two or more storage systems, according to some embodiments of the present disclosure. [0020] FIG. 10 illustrates a flowchart depicting an additional example method for establishing a synchronous replication relationship between two or more storage systems, according to some embodiments of the present disclosure. [0021] FIG. 1 illustrates a flowchart showing an example method for servicing I/O operations directed to data sets that are synchronized across multiple storage systems, in accordance with some embodiments of the present disclosure. [0022] FIG. 10 illustrates a flowchart showing an additional example method for servicing I/O operations directed to data sets that are synchronized across multiple storage systems, in accordance with some embodiments of the present disclosure. [0023] FIG. 10 illustrates a flowchart showing an additional example method for servicing I/O operations directed to data sets that are synchronized across multiple storage systems, in accordance with some embodiments of the present disclosure. [0024] FIG. 10 illustrates a flowchart depicting an additional example method for servicing I/O operations directed to data sets synchronized across multiple storage systems, in accordance with some embodiments of the present disclosure. [0025] FIG. 1 illustrates a flowchart showing an example method for brokering between storage systems that synchronously replicate data sets, according to some embodiments of the present disclosure. [0026] FIG. 1 illustrates a flowchart showing an example method for brokering between storage systems that synchronously replicate data sets, according to some embodiments of the present disclosure. [0027] FIG. 1 illustrates a flowchart showing an example method for brokering between storage systems that synchronously replicate data sets, according to some embodiments of the present disclosure. [0028] FIG. 1 illustrates a flowchart showing an example method for recovery for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0029] FIG. 1 illustrates a flowchart showing an example method for recovery for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0030] FIG. 2 illustrates a flowchart showing an example method for recovery for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0031] FIG. 1 illustrates a flowchart showing an example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0032] FIG. 10 illustrates a flowchart showing an additional example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0033] FIG. 10 illustrates a flowchart showing an additional example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0034] FIG. 10 illustrates a flowchart showing an additional example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0035] FIG. 10 illustrates a flowchart showing an additional example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0036] FIG. 10 illustrates a flowchart showing an additional example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0037] FIG. 10 illustrates a flowchart showing an additional example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0038] FIG. 10 illustrates a flowchart showing an additional example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0039] FIG. 10 illustrates a flowchart showing an additional example method for resynchronization for a storage system that synchronously replicates data sets, according to some embodiments of the present disclosure. [0040] FIG. 1 illustrates a flowchart showing an example method for managing connectivity to a synchronously replicated storage system according to some embodiments of the present disclosure. [0041] FIG. 10 illustrates a flowchart showing an additional example method for managing connectivity to a synchronously replicated storage system in accordance with some embodiments of the present disclosure. [0042] FIG. 10 illustrates a flowchart showing an additional example method for managing connectivity to a synchronously replicated storage system in accordance with some embodiments of the present disclosure. [0043] FIG. 10 illustrates a flowchart showing an additional example method for managing connectivity to a synchronously replicated storage system in accordance with some embodiments of the present disclosure. [0044] FIG. 10 illustrates a flowchart showing an additional example method for managing connectivity to a synchronously replicated storage system in accordance with some embodiments of the present disclosure. [0045] FIG. 10 illustrates a flowchart depicting an additional example method for managing connectivity to a synchronously replicated storage system in accordance with some embodiments of the present disclosure. [0046] FIG. 10 illustrates a flowchart showing an additional example method for managing connectivity to a synchronously replicated storage system in accordance with some embodiments of the present disclosure. [0047] FIG. 10 illustrates a flowchart showing an example method for automated storage system configuration for intermediary services according to some embodiments of the present disclosure. [0048] FIG. 10 illustrates a flowchart showing an example method for automated storage system configuration for intermediary services according to some embodiments of the present disclosure. [0049] FIG. 10 illustrates a flowchart showing an example method for automated storage system configuration for intermediary services according to some embodiments of the present disclosure. [0050] FIG. 10 illustrates a flowchart showing an example method for automated storage system configuration for intermediary services according to some embodiments of the present disclosure. [0051] Figure 1 illustrates a diagram of a metadata representation that may be implemented as a structured collection of metadata objects that together may represent a logical volume or portion of a logical volume of storage data, in accordance with some embodiments of the present disclosure. [0052] FIG. 10 illustrates a flowchart showing an example method for synchronizing metadata between storage systems that synchronously replicate data sets according to some embodiments of the present disclosure. [0053] FIG. 10 illustrates a flowchart showing an example method for synchronizing metadata between storage systems that synchronously replicate data sets according to some embodiments of the present disclosure. [0054] FIG. 10 illustrates a flowchart showing an example method for determining active membership among storage systems that synchronously replicate data sets according to some embodiments of the present disclosure. [0055] FIG. 10 illustrates a flowchart showing an example method for determining active membership among storage systems that synchronously replicate data sets according to some embodiments of the present disclosure. [0056] FIG. 10 illustrates a flowchart showing an example method for determining active membership among storage systems that synchronously replicate data sets according to some embodiments of the present disclosure. [0057] FIG. 10 illustrates a flowchart showing an example method for determining active membership among storage systems that synchronously replicate data sets according to some embodiments of the present disclosure. [0058] FIG. 10 illustrates a flowchart showing an example method for determining active membership among storage systems that synchronously replicate data sets according to some embodiments of the present disclosure. [0059] FIG. 10 illustrates a flowchart showing an example method for synchronizing metadata between storage systems that synchronously replicate data sets according to some embodiments of the present disclosure.

[0060] Example methods, apparatus, and articles of manufacture for synchronously replicating data sets and other managed objects to a cloud-based storage system in accordance with embodiments of the present disclosure are described with reference to the accompanying drawings, beginning with FIG. 1A. FIG. 1A illustrates an example system for data storage according to some embodiments. System 100 (also referred to herein as a "storage system") includes a number of elements for purposes of illustration rather than limitation. Note that system 100 may include the same elements, more elements, or fewer elements configured in the same or different ways in other embodiments.

[0061] System 100 includes several computing devices 164A-164B. For example, computing devices (also referred to herein as "client devices") may be implemented as servers, workstations, personal computers, notebooks, etc. in a data center. Computing devices 164A-164B may be coupled for data communication to one or more storage arrays 102A-102B through a storage area network ("SAN") 158 or a local area network ("LAN") 160.

[0062] SAN 158 may be implemented with a variety of data communication fabrics, devices, and protocols. For example, fabrics for SAN 158 may include Fibre Channel, Ethernet, InfiniBand, Serial Attached Small Computer System Interface ("SAS"), etc. Data communication protocols for use with SAN 158 may include Advanced Technology Attachment ("ATA"), Fibre Channel Protocol, Small Computer System Interface ("SCSI"), Internet Small Computer System Interface ("iSCSI"), HyperSCSI, Non-Volatile Memory Express ("NVMe") over Fabrics, etc. Note that SAN 158 is provided for purposes of illustration and not limitation. Other data communication couplings may be implemented between computing devices 164A-164B and storage arrays 102A-102B.

[0063] LAN 160 may also be implemented with a variety of fabrics, devices, and protocols. For example, fabrics for LAN 160 may include Ethernet (802.3), wireless (802.11), etc. Data communication protocols for use in LAN 160 may include Transmission Control Protocol ("TCP"), User Datagram Protocol ("UDP"), Internet Protocol ("IP"), Hypertext Transfer Protocol ("HTTP"), Wireless Access Protocol ("WAP"), Handheld Device Transfer Protocol ("HDTP"), Session Initiation Protocol ("SIP"), Real Time Protocol ("RTP"), etc.

[0064] Storage arrays 102A-102B may provide persistent data storage for computing devices 164A-164B. In embodiments, storage array 102A may be included in a chassis (not shown) and storage array 102B may be included in another chassis (not shown). Storage arrays 102A and 102B may include one or more storage array controllers 110 (also referred to herein as "controllers"). Storage array controller 110 may be implemented as an automated computing module including computer hardware, computer software, or a combination of computer hardware and software. In some embodiments, storage array controller 110 may be configured to perform a variety of storage tasks. Storage tasks may include writing data received from computing devices 164A-164B to storage arrays 102A-102B, erasing data from storage arrays 102A-102B, retrieving data from storage arrays 102A-102B and providing the data to computing devices 164A-164B, monitoring and reporting disk utilization and performance, performing redundancy operations such as RAID or RAID-like data redundancy operations, compressing data, encrypting data, etc.

[0065] Storage array controller 110 may be implemented in a variety of ways, including as a field programmable gate array ("FPGA"), a programmable logic chip ("PLC"), an application specific integrated circuit ("ASIC"), a system on a chip ("SOC"), or any computing device including discrete components such as a processing unit, a central processing unit, computer memory, or various adapters. Storage array controller 110 may include, for example, a data communications adapter configured to support communications via SAN 158 or LAN 160. In some embodiments, storage array controller 110 may be independently coupled to LAN 160. In embodiments, storage array controller 110 may include an I/O controller or the like that couples storage array controller 110 for data communications to persistent storage resources 170A-170B (also referred to herein as "storage resources") through a midplane (not shown). Persistent storage resources 170A-170B may include any number of storage drives 171A-171F (also referred to herein as "storage devices") and any number of non-volatile random access memory (NVRAM) devices (not shown).

[0066] In some embodiments, the NVRAM devices of persistent storage resources 170A-170B may be configured to receive data to be stored on storage drives 171A-171F from storage array controller 110. In some examples, the data may originate from computing devices 164A-164B. In some examples, writing data to the NVRAM devices may be performed more quickly than writing data directly to storage drives 171A-171F. In embodiments, storage array controller 110 may be configured to utilize the NVRAM devices as a quickly accessible buffer for data known to be written to storage drives 171A-171F. The latency of write requests using the NVRAM devices as a buffer may be improved compared to systems in which storage array controller 110 writes data directly to storage drives 171A-171F. In some embodiments, the NVRAM devices may be implemented with computer memory in the form of high-bandwidth, low-latency RAM. NVRAM devices are referred to as "non-volatile" because they may receive or include an intrinsic power source that maintains the state of the RAM after a loss of main power to the NVRAM device. Such a power source may be a battery, one or more capacitors, etc. In response to a power loss, the NVRAM device may be configured to write the contents of the RAM to persistent storage, such as storage drives 171A-171F.

[0067] In embodiments, storage drives 171A-171F may refer to any device configured to persistently record data, where "persistently" or "persistent" refers to the device's ability to maintain recorded data after a loss of power. In some embodiments, storage drives 171A-171F may correspond to non-disk storage media. For example, storage drives 171A-171F may be one or more solid-state drives ("SSDs"), flash memory-based storage, any type of solid-state non-volatile memory, or any other type of non-mechanical storage device. In other embodiments, storage drives 171A-171F may include mechanical or rotating hard disks, such as hard disk drives ("HDDs").

[0068] In some embodiments, the storage array controller 110 may be configured to remove device management responsibilities from the storage drives 171A-171F in the storage arrays 102A-102B. For example, the storage array controller 110 may manage control information that may describe the state of one or more memory blocks of the storage drives 171A-171F. The control information may indicate, for example, that a particular memory block has failed and should no longer be written to, that a particular memory block contains boot code for the storage array controller 110, the number of program-erase ("P/E") cycles that have been performed on a particular memory block, the age of the data stored in a particular memory block, the type of data stored in a particular memory block, etc. In some embodiments, the control information may be stored with the associated memory block as metadata. In other embodiments, the control information for the storage drives 171A-171F may be stored in one or more specific memory blocks of the storage drives 171A-171F selected by the storage array controller 110. The selected memory blocks may be tagged with an identifier indicating that the selected memory blocks contain control information. The identifier may be utilized by storage array controller 110 in conjunction with storage drives 171A-171F to quickly identify memory blocks containing control information. For example, storage controller 110 may issue a command to locate memory blocks containing control information. Note that because the control information is quite large, portions of the control information may be stored in multiple locations; the control information may be stored in multiple locations, for example, for redundancy, or the control information may otherwise be distributed across multiple memory blocks on storage drives 171A-171F.

[0069] In embodiments, storage array controller 110 may remove device management responsibility from storage drives 171A-171F of storage arrays 102A-102B by retrieving, from storage drives 171A-171F, control information describing the state of one or more memory blocks of storage drives 171A-171F. Retrieving the control information from storage drives 171A-171F may be performed, for example, by storage array controller 110 querying storage drives 171A-171F about the location of the control information for a particular storage drive 171A-171F. Storage drives 171A-171F may be configured to execute instructions that enable storage drives 171A-171F to identify the location of the control information. The instructions may be executed by a controller (not shown) associated with or otherwise located at storage drives 171A-171F, and may cause storage drives 171A-171F to scan a portion of each memory block to identify memory blocks that store control information for storage drives 171A-171F. Storage drives 171A-171F may respond by sending a response message to storage array controller 110 that includes the location of the control information for storage drives 171A-171F. In response to receiving the response message, storage array controller 110 may issue a request to read the data stored at the address associated with the location of the control information for storage drives 171A-171F.

[0070] In other embodiments, storage array controller 110 may further remove device management responsibilities from storage drives 171A-171F by performing storage drive management operations in response to receiving the control information. Storage drive management operations may include, for example, operations normally performed by storage drives 171A-171F (e.g., a controller (not shown) associated with a particular storage drive 171A-171F). Storage drive management operations may include, for example, ensuring that data is not written to failed memory blocks within storage drives 171A-171F, ensuring that data is written to memory blocks within storage drives 171A-171F so that an appropriate wear rate is achieved, etc.

[0071] In embodiments, storage arrays 102A-102B may implement two or more storage array controllers 110. For example, storage array 102A may include storage array controller 110A and storage array controller 110B. At a given instance, a single storage array controller 110 (e.g., storage array controller 110A) in storage system 100 may be designated with primary status (also referred to herein as the "primary controller"), and the other storage array controller 110 (e.g., storage array controller 110A) may be designated with secondary status (also referred to herein as the "secondary controller"). The primary controller may have certain rights, such as the permission to modify data on persistent storage resources 170A-170B (e.g., writing data to persistent storage resources 170A-170B). At least some of the rights of the primary controller may supersede the rights of the secondary controller. For example, a secondary controller may not have permission to modify data on persistent storage resources 170A-170B when the primary controller has the right. The status of storage array controllers 110 may change. For example, storage array controller 110A may be designated with secondary status, and storage array controller 110B may be designated with primary status.

[0072] In some embodiments, a primary controller, such as storage array controller 110A, may serve as the primary controller for one or more storage arrays 102A-102B, and a second controller, such as storage array controller 110B, may serve as a secondary controller for one or more storage arrays 102A-102B. For example, storage array controller 110A may be the primary controller for storage arrays 102A and 102B, and storage array controller 110B may be the secondary controller for storage arrays 102A and 102B. In some embodiments, storage controllers 110C and 110D (also referred to as "storage processing modules") may have neither primary nor secondary status. Storage array controllers 110C and 110D, implemented as storage processing modules, may serve as a communication interface between the primary and secondary controllers (e.g., storage array controllers 110A and 110B, respectively) and storage array 102B. For example, storage array controller 110A of storage array 102A may send a write request to storage array 102B via SAN 158. The write request may be received by both storage array controllers 110C and 110D of storage array 102B. Storage array controllers 110C and 110D facilitate the communication, e.g., sending the write request to the appropriate storage drives 171A-171F. Note that in some embodiments, storage processing modules may be used to increase the number of storage drives controlled by the primary and secondary controllers.

[0073] In an embodiment, storage array controller 110 is communicatively coupled to one or more storage drives 171A-171F and to one or more NVRAM devices (not shown) included as part of storage arrays 102A-102B via a midplane (not shown). Storage array controller 110 may be coupled to the midplane via one or more data communication links, and the midplane may be coupled to storage drives 171A-171F and the NVRAM devices via one or more data communication links. The data communication links described herein are collectively represented by data communication links 108A-108D and may include, for example, Peripheral Component Interconnect Express ("PCIe") buses.

[0074] FIG. 1B illustrates an example system for data storage according to some embodiments. The storage array controller 101 illustrated in FIG. 1B may be similar to the storage array controller 110 described with respect to FIG. 1A. In one example, the storage array controller 101 may be similar to storage array controller 110A or storage array controller 110B. The storage array controller 101 includes multiple elements for purposes of explanation rather than limitation. Note that the storage array controller 101 may include the same elements, more elements, or fewer elements configured similarly or differently in other embodiments. Note that elements of FIG. 1A may be included below to help illustrate features of the storage array controller 101.

[0075] The storage array controller 101 may include one or more processing units 104 and random access memory ("RAM") 111. The processing unit 104 (or controller 101) represents one or more general-purpose processing units, such as, for example, a microprocessor, a central processing unit, or the like. More specifically, the processing unit 104 (or controller 101) may be a multiple instruction set computing ("CISC") microprocessor, a reduced instruction set computing ("RISC") microprocessor, a very long instruction word ("VLIW") microprocessor, or a processor implementing other instruction sets or a processor implementing a combination of instruction sets. The processing unit 104 (or controller 101) may also be one or more special-purpose processing units, such as, for example, an application specific integrated circuit ("ASIC"), a field programmable gate array ("FPGA"), a digital signal processor ("DSP"), a network processor, or the like.

[0076] Processing unit 104 may be connected to RAM 111 via data communication link 106, which may be implemented as a high-speed memory bus such as a double data rate 4 ("DDR4") bus. Stored in RAM 111 is operating system 112. In some implementations, instructions 113 are stored in RAM 111. Instructions 113 may include computer program instructions for performing operations on a direct-mapped flash storage system. In one embodiment, a direct-mapped flash storage system is a flash storage system that directly addresses data blocks within a flash drive, without address translation, and is executed by the flash drive's storage controller.

[0077] In an embodiment, the storage array controller 101 includes one or more host bus adapters 103A-103C coupled to the processing unit 104 via data communication links 105A-105C. In an embodiment, the host bus adapters 103A-103C may be computer hardware that connects a host system (e.g., a storage array controller) to other networks and storage arrays. In some examples, the host bus adapters 103A-103C may be Fibre Channel adapters that allow the storage array controller 101 to connect to a SAN, Ethernet adapters that allow the storage array controller 101 to connect to a LAN, etc. The host bus adapters 103A-103C may be coupled to the processing unit 104 via data communication links 105A-105C, such as a PCIe bus.

[0078] In an embodiment, storage array controller 101 may include a host bus adapter 114 coupled to an expander 115. Expander 115 may be used to attach a host system to a larger number of storage drives. Expander 115 may be, for example, a SAS expander utilized to allow host adapter 114 to attach storage drives in an embodiment in which host bus adapter 114 is implemented as a SAS controller.

[0079] In an embodiment, storage array controller 101 may include a switch 116 coupled to processing unit 104 via data communication link 109. Switch 116 may be a computer hardware device that creates multiple endpoints from a single endpoint, thereby allowing multiple devices to share a single endpoint. Switch 116 may be, for example, a PCIe switch coupled to a PCIe bus (e.g., data bus link 109) and presenting multiple PCIe connection points at a midplane.

[0080] In embodiments, storage array controller 101 includes data communication link 107 for coupling storage array controller 101 to other storage array controllers. In some examples, data communication link 107 may be a Quick Path Interconnect (QPI) interconnect.

[0081] A conventional storage system using conventional flash drives may implement processes throughout the flash drives that are part of the conventional storage system. For example, higher-level processes in the storage system may initiate and control processes throughout the flash drives. However, flash drives in a conventional storage system may include their own storage controller that also executes processes. Thus, in the case of a conventional storage system, both higher-level processes (e.g., initiated by the storage system) and lower-level processes (e.g., initiated by the storage controller of the storage system) may be executed.

[0082] To address various deficiencies of conventional storage systems, operations may be performed by higher-level processes rather than by lower-level processes. For example, a flash storage system may include a flash drive that does not include a storage controller to provide the processes. Thus, the flash storage system's operating system itself may initiate and control the processes. This is accomplished by direct mapping flash storage systems that address data blocks directly within the flash drive, which may be performed by the flash drive's storage controller without address translation.

[0083] The operating system of the flash storage system may identify and maintain a list of allocation units across multiple flash drives in the flash storage system. An allocation unit may be a whole erase block or multiple erase blocks. The operating system may maintain a map or address ranges that directly map addresses to erase blocks on the flash drives of the flash storage system.

[0084] Direct mapping to erase blocks on a flash drive may be used to rewrite data and erase data. For example, operations may be performed on one or more allocation units containing first data and second data, where the first data is retained and the second data is no longer in use by the flash storage system. The operating system may initiate a process to write the first data to a new location in another allocation unit, erase the second data, and mark the allocation unit as available for use for subsequent data. In this manner, the process may be performed solely by the higher-level operating system of the flash storage system, without additional lower-level processes being performed by the flash drive's controller.

[0085] Advantages of processes being performed solely by the flash storage system's operating system include increased reliability of the flash drives in the flash storage system, since unnecessary or redundant write operations are not performed during the process. One possible point of novelty here is the concept of initiating and controlling the process in the flash storage system's operating system. Furthermore, the process can be controlled by the operating system across multiple flash drives. This is in contrast to processes being performed by the flash drive's storage controller.

[0086] A storage system may consist of two storage array controllers that share a set of drives for failover purposes, or the storage system may consist of a single storage array controller that provides storage services utilizing multiple drives, or the storage system may consist of a distributed network of storage array controllers, each with some number of drives or some amount of flash storage, that cooperate to provide a complete storage service and cooperate with respect to various aspects of the storage service, including storage allocation and garbage collection.

[0087] Figure 1C illustrates a third example system 117 for data storage according to some embodiments. System 117 (also referred to herein as a "storage system") includes a number of elements for purposes of illustration rather than limitation. Note that system 117 may include the same elements, more elements, or fewer elements configured similarly or differently in other embodiments.

[0088] In one embodiment, system 117 includes dual peripheral component interconnect ("PCI") flash storage devices 118 with separately addressable high-speed write storage. System 117 may include storage controller 119. In one embodiment, storage controller 119 may be a CPU, ASIC, FPGA, or any other circuitry that may implement the necessary control structures in accordance with the present disclosure. In one embodiment, system 117 includes flash memory devices (e.g., including flash memory devices 120a through 120n) operatively coupled to various channels of storage device controller 119. Flash memory devices 120a through 120n may be presented to controller 119 as addressable collections of flash pages, erase blocks, and/or control elements sufficient to allow storage device controller 119 to program and retrieve various aspects of the flash. In one embodiment, storage device controller 119 may perform operations on flash memory devices 120A-120N, including storing and retrieving data content of pages, allocating and erasing any blocks, tracking statistics related to the use and reuse of flash memory pages, erase blocks, and cells, tracking and predicting error codes and failures in flash memory, controlling voltage levels associated with programming, and retrieving content such as flash cells.

[0089] In one embodiment, system 117 may include RAM 121 for storing separately addressable high-speed write data. In one embodiment, RAM 121 may be one or more separate, discrete components. In another embodiment, RAM 121 may be integrated into storage device controller 119 or multiple storage device controllers. RAM 121 may be utilized for other purposes, such as temporary program memory for a processing unit (e.g., a CPU) within storage device controller 119.

[0090] In one embodiment, system 119 may include a stored energy device 122, such as a rechargeable battery or a capacitor. Stored energy device 122 may store enough energy to power some amount of RAM (e.g., RAM 121), some amount of flash memory (e.g., flash memories 120a through 120n), and storage device controller 119 for a time sufficient to write the contents of the RAM to flash memory. In one embodiment, storage device controller 119 may write the contents of RAM to flash memory when the storage device controller detects a loss of external power.

[0091] In one embodiment, system 117 includes two data communication links 123a, 123b. In one embodiment, data communication links 123a, 123b may be PCI interfaces. In another embodiment, data communication links 123a, 123b may be based on other communication standards (e.g., HyperTransport, InfiniBand, etc.). Data communication links 123a, 123b may be based on the Non-Volatile Memory Express ("NVMe") or NVMe over Fabric ("NVMf") specifications, which enable external connections from other components of storage system 117 to storage device controller 119. Note that for convenience, data communication links may be referred to herein synonymously as PCI buses.

[0092] System 117 may also include an external power source (not shown), which may be provided via one or both data communication links 123a, 123b, or may be provided separately. An alternative embodiment includes a separate flash memory (not shown) dedicated for use in storing the contents of RAM 121. Storage device controller 119 may present a logical device on the PCI bus, which may include an addressable fast-write logical device, or a separate portion of the logical address space of storage device 118, which may be presented as PCI memory or as persistent storage. In one embodiment, operations to store into the device are directed into RAM 121. In the event of a power outage, storage device controller 119 may write stored content associated with the addressable fast-write logical storage to flash memory (e.g., flash memories 120a through 120n) for long-term persistent storage.

[0093] In one embodiment, the logical device may include some representation of some or all of the contents of flash memory devices 120a-120n that allows a storage system (e.g., storage system 117) including storage device 118 to directly address flash memory pages and reprogram erase blocks directly from storage system components external to the storage device through a PCI bus. The representation may also allow one or more of the external components to control and retrieve other aspects of the flash memory, including some or all of the following: tracking statistics related to the use and reclamation of flash memory pages, erase blocks, and cells across all flash memory devices; tracking and predicting error codes and failures within and across flash memory devices; controlling voltage levels associated with programming; and retrieving content such as flash cells.

[0094] In one embodiment, stored energy device 122 may be sufficient to ensure completion of ongoing operations on flash memory devices 107a-120n, and stored energy device 122 may provide power to storage device controller 119 and associated flash memory devices (e.g., 120a-120n) for those operations and for fast write RAM storage to flash memory. Stored energy device 122 may also be used to store accumulated statistics and other parameters kept and tracked by flash memory devices 120a-120n and/or storage device controller 119. A separate capacitor or stored energy device (e.g., a smaller capacitor near or embedded in the flash memory device itself) may be used for some or all of the operations described herein.

[0095] Various approaches may be used to track and optimize the life of the stored energy components, such as adjusting the voltage level over time or partially discharging the stored energy device 122 and measuring the corresponding discharge characteristics. If the available energy decreases over time, the effective available capacity of the addressable fast-write storage may be reduced to ensure that it can be safely written based on the currently available stored energy.

[0096] Figure 1D illustrates a third example system 124 for data storage in accordance with some embodiments. In one embodiment, system 124 includes storage controllers 125a and 125b. In one embodiment, storage controllers 125a and 125b are operably coupled to dual PCI storage devices 119a, 119b, and 119c, 119d, respectively. Storage controllers 125a and 125b may be operably coupled to any number of host computers 127a-127n (e.g., via storage network 130).

[0097] In one embodiment, two storage controllers (e.g., 125a and 125b) provide storage services, such as a block storage array (e.g., SCSI), a file server, an object server, a database, or a data analysis service. Storage controllers 125a, 125b may provide services to host computers 127a-127n outside of storage system 124 through any number of network interfaces (e.g., 126a-126d). Storage controllers 125a, 125b may provide integrated services or applications entirely within storage system 124, forming a converged storage and computing system. Storage controllers 125a, 125b may utilize high-speed write memory within or across storage devices 119a-119d to record ongoing operations to ensure that operations are not lost in the event of a power outage, removal of a storage controller, a shutdown of a storage controller or storage system, or any failure of one or more software or hardware components within storage system 124.

[0098] In one embodiment, controllers 125a, 125b act as PCI masters for one or the other PCI bus 128a, 128b. In another embodiment, 128a and 128b may be based on other communication standards (e.g., HyperTransport, InfiniBand, etc.). Other storage system embodiments may operate storage controllers 125a, 125b as multi-masters for PCI bus 128a, 128b. Alternatively, a PCI/NVMe/NVMf switching infrastructure or fabric may connect multiple storage controllers. Some storage system embodiments may allow storage devices to communicate directly with each other rather than communicating only with the storage controller. In one embodiment, storage device controller 119a may be operable under direction from storage controller 125a to combine data stored in flash memory devices and transfer data from data stored in RAM (e.g., RAM 121 of FIG. 1C). For example, a recomputed version of the RAM contents may be transferred after the storage controller determines that operations are fully committed across the storage system, or when fast write memory for a device reaches a certain usage capacity, or after a certain amount of time, to improve data safety or to free up addressable fast write capacity for reuse. This mechanism may be used, for example, to avoid a second transfer on a bus (e.g., 128a, 128b) from storage controller 125a, 125b. In one embodiment, the recomputation may include compressing data, attaching indexing or other metadata, combining multiple data segments together, performing erasure code calculations, etc.

[0099] In one embodiment, under direction from storage controller 125a, 125b, storage device controller 119a, 119b may operate to calculate and transfer data from data stored in RAM (e.g., RAM 121 in FIG. 1C) to other storage devices without the involvement of storage controller 125a, 125b. This operation may be used to mirror data stored from one controller 125a to another controller 125b, or it may be used to offload compression, data consolidation, and/or erasure coding calculations and transfers to storage devices to reduce the load on PCI bus 128a, 128b from storage controller or storage controller interface 129a, 129b.

[0100] The storage device controller 119 may include mechanisms for implementing high availability primitives for use by other portions of the storage system external to the dual PCI storage device 118. For example, reservation or exclusion primitives may be provided so that in a storage system having two storage controllers providing highly available storage services, one storage controller may prevent the other storage controller from accessing or continuing to access the storage device. This could be used, for example, when one controller detects that the other controller is not functioning properly, or when the interconnect between the two storage controllers may itself not be functioning properly.

[0101] In one embodiment, a storage system for use with a dual PCI direct mapped storage device having separately addressable fast write storage includes a system that manages erase blocks or groups of erase blocks as allocation units for storing data on behalf of a storage service, for storing metadata associated with a storage service (e.g., indexes, logs, etc.), or for proper management of the storage system itself. Flash pages, which may be several kilobytes in size, may be written as data arrives or once the storage system has persisted the data for a long interval of time (e.g., beyond a defined threshold of time). To commit data more quickly or to reduce the number of writes to the flash memory device, the storage controller may first write the data to separately addressable fast write storage on one or more storage devices.

[0102] In one embodiment, storage controllers 125a, 125b may initiate the use of erase blocks within and across storage devices (e.g., 118) according to the age and expected remaining life of the storage device, or based on other statistics. Storage controllers 125a, 125b may initiate garbage collection and data migration data between storage devices according to pages that are no longer needed, as well as to manage the lifespan of flash pages and erase blocks and manage overall system performance.

[0103] In one embodiment, storage system 124 may utilize mirroring and/or erasure coding schemes as part of storing data to addressable fast-write storage and/or as part of writing data to allocation units associated with erase blocks. Erasure codes may be used within erase blocks or allocation units, or across flash memory devices on a single storage device, as well as across storage devices, to provide redundancy against single or multiple storage device failures or to protect against internal corruption of flash memory pages resulting from flash memory operations or from degradation of flash memory cells. Various levels of mirroring and erasure coding may be used to recover from multiple types of failures, either separately or in combination.

2A through 2G illustrate a storage cluster that stores user data, such as user data originating from one or more user or client systems or other sources external to the storage cluster. The storage cluster uses erasure coding and redundant copies of metadata to distribute user data across storage nodes housed within a chassis or across multiple chassis. Erasure coding refers to a method of protecting or reconstructing data where the data is stored across a collection of different locations, such as disks, storage nodes, or geographic locations. Flash memory is one type of solid-state memory that may be integrated with embodiments, although embodiments may be extended to other storage media, including other types of solid-state memory or non-solid-state memory. Storage location and workload control are distributed across storage locations in a clustered peer-to-peer system. Tasks such as brokering communications between the various storage nodes, detecting when a storage node becomes unavailable, and balancing I/O (input/output) across the various storage nodes are all handled on a distributed basis. Data is laid out or distributed across multiple storage nodes in data fragments or stripes, which in some embodiments supports data recovery. Data ownership is reassigned within the cluster regardless of input and output patterns. This architecture, described in more detail below, allows storage nodes within a cluster to fail while the system remains operational, since data can be reconstructed from other storage nodes and therefore remains available for input and output operations. In various embodiments, storage nodes may be referred to as cluster nodes, blades, or servers.

[0105] A storage cluster may be contained within a chassis, i.e., an enclosure that houses one or more storage nodes. Contained within the chassis are mechanisms for providing power to each storage node, such as a power distribution bus, and communication mechanisms, such as a communication bus, that enable communication between the storage nodes. A storage cluster, according to some embodiments, can run as an independent system in a single location. In one embodiment, a chassis includes at least two instances of both a power distribution and a communication bus, which can be independently enabled or disabled. The internal communication bus may be an Ethernet bus, although other technologies, such as PCIe, InfiniBand, and others, are equally suitable. The chassis provides ports for an external communication bus to enable communication between multiple chassis and with client systems, either directly or through a switch. External communication may use technologies such as Ethernet, InfiniBand, Fibre Channel, etc. In some embodiments, the external communication bus uses different communication bus technologies for inter-chassis communication and client communication. When a switch is deployed within or between chassis, the switch may perform the function of translating between multiple protocols or technologies. When multiple chassis are connected to define a storage cluster, the storage cluster may be accessed by clients using a proprietary interface or a standard interface, such as Network File System (NFS), Common Internet File System (CIFS), Small Computer System Interface (SCSI), or Hypertext Transfer Protocol (HTTP). Translation from the client protocol may occur at the switch, the chassis external communication bus, or within each storage node. In some embodiments, multiple chassis may be coupled or connected to each other through an aggregator switch. Some and/or all of the coupled or connected chassis may be designated as a storage cluster. As mentioned above, each chassis may have multiple blades, and each blade has a media access control (MAC) address, but in some embodiments the storage cluster is presented to the external network as having a single cluster IP address and a single MAC address.

[0106] Each storage node may be one or more storage servers, each connected to one or more non-volatile solid-state memory units, sometimes referred to as storage units or storage devices. One embodiment includes a single storage server for each storage node and between one and eight non-volatile solid-state memory units, although this example is not intended to be limiting. The storage server may include a processor, DRAM, and interfaces for internal communication buses and power distribution, respectively. In some embodiments, inside the storage node, the interfaces and storage units share a communication bus, such as PCI Express. The non-volatile solid-state memory units may directly access the internal communication bus interface through the storage node communication bus or may require the storage node to access the bus interface. The non-volatile solid-state memory units include an embedded CPU, a solid-state storage controller, and, in some embodiments, a large amount of solid-state mass storage, e.g., between 2 and 32 terabytes ("TB"). An embedded volatile storage medium, such as DRAM, and an energy storage device are included in the non-volatile solid-state memory units. In some embodiments, the energy storage device is a capacitor, supercapacitor, or battery that allows for the transfer of a subset of the DRAM contents to a stable storage medium upon power loss. In some embodiments, the non-volatile solid-state memory unit is constructed with storage-class memory, such as phase-change memory or magnetoresistive random access memory ("MRAM"), which replaces DRAM and allows for reduced power hold-up apparatus.

[0107] One of the many features of storage nodes and non-volatile solid-state storage is the ability to proactively rebuild data within a storage cluster. The storage nodes and non-volatile solid-state storage can determine when a storage node or non-volatile solid-state storage within a storage cluster is unreachable, regardless of whether there is an attempt to read the data associated with that storage node or non-volatile solid-state storage. The storage nodes and non-volatile solid-state storage then cooperate to at least partially recover and rebuild the data in a new location. This constitutes proactive rebuilding in that the system rebuilds data without waiting until the data is needed for a read access initiated by a client system utilizing the storage cluster. These and additional details of the storage memory and its operation are described below.

2A is a perspective view of a storage cluster 161 having multiple storage nodes 150 and internal solid-state memory coupled to each storage node to provide network-attached storage or a storage area network, according to some embodiments. Network-attached storage, storage area network, storage cluster, or other storage memory may include one or more storage clusters 161, each having one or more storage nodes 150, with a flexible and reconfigurable configuration of both physical components and the amount of storage memory provided thereby. The storage cluster 161 is designed to fit into a rack, and one or more racks may be set up and populated as desired for the storage memory. The storage cluster 161 includes a chassis 138 having multiple slots 142. It should be understood that the chassis 138 may also be referred to as a case, enclosure, or rack unit. While other slot counts are readily devised, in one embodiment, the chassis 138 has 14 slots 142. For example, some embodiments have four slots, eight slots, 16 slots, 32 slots, or other suitable numbers of slots. In some embodiments, each slot 142 may house one storage node 150. The chassis 138 includes a flap 148 that can be utilized to mount the chassis 138 on a rack. The fans 144 provide air circulation for cooling the storage nodes 150 and their components, although other cooling components could be used, or embodiments without cooling components could be devised. The switch fabric 146 couples the storage nodes 150 in the chassis 138 to each other and to a network for communication to memory. In the embodiment shown herein, the slots 142 to the left of the switch fabric 146 and fans 144 are shown occupied by storage nodes 150, while the slots 142 to the right of the switch fabric 146 and fans 144 are empty and available for insertion of storage nodes 150 for illustrative purposes. This configuration is exemplary, and one or more storage nodes 150 could occupy the slots 142 in a variety of additional configurations. The configuration of storage nodes need not be contiguous or adjacent in some embodiments. Storage node 150 is hot-pluggable, meaning that storage node 150 can be inserted into or removed from slot 142 in chassis 138 without shutting down or powering down the system. When storage node 150 is inserted or removed from slot 142, the system recognizes the change and automatically reconfigures to accommodate the change. In some embodiments, the reconfiguration includes restoring redundancy and/or rebalancing data or load.

[0109] Each storage node 150 may include multiple components. While other attachments and/or components may be used in additional embodiments, in the embodiment shown, storage node 150 includes a CPU 156, i.e., a printed circuit board 159 mounted with a processor, memory 154 coupled to CPU 156, and non-volatile solid-state storage 152 coupled to CU 156. Memory 154 contains instructions executed by CPU 156 and/or data affected by CPU 156. As described further below, non-volatile solid-state storage 152 may include flash, or in additional embodiments, other types of solid-state memory.

2A, the storage cluster 161 is scalable, meaning that storage capacity having non-uniform storage sizes is easily added as described above. One or more storage nodes 150 can be plugged in or removed from each chassis, and the storage cluster is self-configuring in some embodiments. The plug-in storage nodes 150, whether installed in the chassis as delivered or as a later add-on, may have different sizes. For example, in one embodiment, the storage nodes 150 may have any multiple of 4 TB, such as 8 TB, 12 TB, 16 TB, 32 TB, etc. In additional embodiments, the storage nodes 150 could have other storage amounts or any multiple of storage capacity. The storage capacity of each storage node 150 is broadcast and influences the determination of how the data is striped. For maximum storage efficiency, embodiments may self-configure in stripes as widely as possible, subject to the given requirement of continued operation with the loss of at most one or at most two non-volatile solid-state storage units 152 or storage nodes 150 in the chassis.

2B is a block diagram illustrating communication interconnects 171A-171F and power distribution bus 172 coupling multiple storage nodes 150. Referring back to FIG. 2A, in some embodiments, communication interconnects 171A-171F may be included in or implemented in switch fabric 146. When multiple storage clusters 161 occupy a rack, in some embodiments, communication interconnects 171A-171F may be included in or implemented in an upper portion of a rack switch. As shown in FIG. 2B, storage cluster 161 is enclosed within a single chassis 138. External port 176 is coupled to storage node 150 through communication interconnects 171A-171F, while external port 174 is coupled directly to the storage node. External power port 178 is coupled to power distribution bus 172. Storage node 150 may include varying amounts and capacities of non-volatile solid-state storage 152, as described with respect to FIG. 2A . Additionally, one or more storage nodes may be compute-only storage nodes, as shown in FIG. 2B . Authority 168 is embodied on non-volatile solid-state storage 152, for example, as a list or other data structure stored in memory. In some embodiments, authority 168 is stored in non-volatile solid-state storage 152 and supported by software running on a controller or other processor of non-volatile solid-state storage 152. In additional embodiments, authority 168 is embodied on storage node 150, for example, as a list or other data structure stored in memory 154, and supported by software running on CPU 156 of storage node 150. Authority 168, in some embodiments, controls how and where data is stored on non-volatile solid-state storage 152. This control helps determine what type of erasure coding scheme is applied to the data and which portions of the data storage node 150 holds. Each authority 168 may be assigned to non-volatile solid-state storage 152. In various embodiments, each authority may control a set of inode numbers, segment numbers, or other data identifiers assigned to data by the file system, by storage node 150, or by non-volatile solid-state storage 152.

[0112] In some embodiments, every piece of data and every piece of metadata has redundancy within the system. Furthermore, every piece of data and every piece of metadata has an owner, sometimes called an authority. If the authority is unreachable, for example due to a storage node failure, there is a succession plan for how to find the data or its metadata. In various embodiments, there are redundant copies of authority 168. In some embodiments, authority 168 has a relationship to storage nodes 150 and non-volatile solid-state storage 152. Each authority 168, which covers a range of data segment numbers or other identifiers of data, may be assigned to a particular non-volatile solid-state storage 152. In some embodiments, authorities 168 for all such ranges are distributed across the non-volatile solid-state storage 152 of the storage cluster. Each storage node 150 has a network port that provides access to the non-volatile solid-state storage(s) 152 of that storage node 150. Data may be stored in segments that are associated with a segment number, which in some embodiments is the indirection for the configuration of a RAID (Redundant Array of Independent Disks) stripe. The assignment and use of authority 168 thus establishes indirection to the data. Indirection, according to some embodiments, may be referred to as the ability to reference data indirectly, in this case via authority 168. A segment identifies a collection of non-volatile solid-state storage 152 and a local identifier within the collection of non-volatile solid-state storage 152 that may contain the data. In some embodiments, the local identifier is an offset into the device and may be reused consecutively by multiple segments. In other embodiments, the local identifier is unique to a particular segment and is never reused. The non-volatile solid-state storage 152 offset is applied to locate data for writing to or reading from non-volatile solid-state storage 152 (in the form of a RAID stripe). Data is striped across multiple units of non-volatile solid-state storage 152, which may include non-volatile solid-state storage 152 with authority 168 for a particular data segment, or may be different from non-volatile solid-state storage 152.

[0113] For example, during data migration or data reconstruction, if there is a change in the location of a particular segment of data, the authority 168 for that data segment should be consulted in that non-volatile solid-state storage 152 or storage node 150 that has that authority 168. To locate a particular piece of data, embodiments calculate a hash value of the data segment or apply an inode number or data segment number. The output of this operation points to the non-volatile solid-state storage 152 that has the authority 168 for that piece of data. In some embodiments, there are two stages to this operation. The first stage is mapping an entity identifier (ID), such as a segment number, inode number, or directory number, to an authority identifier. This mapping may include the calculation of a hash or bitmask, etc. The second stage is mapping the authority identifier to a particular non-volatile solid-state storage 152, which may be done through explicit mapping. The operation is repeatable, so that when the calculation is performed, the result of the calculation repeatably and reliably points to the particular non-volatile solid-state storage 152 with the authority 168. The operation may include a set of reachable storage nodes as input. If the set of reachable non-volatile solid-state units changes, the optimal set changes. In some embodiments, the persistence value is the current allocation (which is always true), and the calculated value is the target allocation toward which the cluster attempts to reconfigure. This calculation may be used to determine the optimal non-volatile solid-state storage 152 for the authority given the set of non-volatile solid-state storages 152 that are reachable and that constitute the same cluster. The calculation also determines an ordered set of peer non-volatile solid-state storages 152 that also record the authority to non-volatile solid-state storage mapping so that an authority can be determined even if the assigned non-volatile solid-state storage is unreachable. If a particular authority 168 is unavailable in some embodiments, a duplicate or alternative authority 168 may be consulted.

2A and 2B, two of the many tasks of the CPU 156 on a storage node 150 are to split write data and reassemble read data. When the system determines that data is to be written, the authority 168 for that data is located as described above. When the segment ID of the data has already been determined, the write request is forwarded from the segment to the non-volatile solid-state storage 152 currently determined to host the determined authority 168. The host CPU 156 of the storage node 150 on which the non-volatile solid-state storage 152 and corresponding authority 168 reside then splits or shards the data and sends the data to the various non-volatile solid-state storages 152. The transmitted data is written as data stripes according to an erasure coding scheme. In some embodiments, the data is requested to be pulled, while in other embodiments, the data is pushed. Conversely, when data is to be read, the authority 168 for the segment ID containing the data is located as described above. The host CPU 156 of the storage node 150 on which the non-volatile solid-state storage 152 and corresponding authority 168 reside requests data from the non-volatile solid-state storage and corresponding storage node pointed to by the authority. In some embodiments, the data is read from flash storage as a data stripe. The host CPU 156 of the storage node 150 then reassembles the read data, corrects any errors (if present) according to an appropriate erasure coding scheme, and transfers the reassembled data to the network. In additional embodiments, some or all of these tasks may be handled within the non-volatile solid-state storage 152. In some embodiments, the segment host requests that data be sent to the storage node 150 by requesting a page from storage and then sending the data to the storage node that made the original request.

[0115] In some systems, such as UNIX-style file systems, data is organized in index nodes or inodes, which specify data structures that represent objects in the file system. An object might be, for example, a file or a directory. Metadata may be associated with an object as attributes such as permission data and creation timestamps, among other attributes. A segment number might be assigned to all or a portion of such an object in the file system. In other systems, data segments are organized with segment numbers assigned elsewhere. For purposes of illustration, the unit of distribution is an entity, which might be a file, directory, or segment. That is, an entity is a unit of data or metadata stored by the storage system. Entities are grouped into collections called authorities. Each authority has an authority owner, which is a storage node that has exclusive rights to update the authority's entities. In other words, storage nodes contain authorities, which in turn contain entities.

[0116] According to some embodiments, a segment is a logical container of data. A segment is an address space between the media address space and a physical flash location; that is, data segment numbers reside within this address space. A segment may also contain metadata that allows data redundancy to be restored (rewritten to a different flash location or device) without the involvement of higher-level software. In one embodiment, the internal format of a segment includes client data and intermediate mappings to determine the location of the data. Each data segment is protected against, for example, memory and other failures, if applicable, by dividing the segment into several data and parity shards. The data and parity shards are distributed, or striped, across the non-volatile solid-state storage 152 coupled to the host CPU 156 (see Figures 2E and 2G) according to an erasure coding scheme. The use of the term segment, in some embodiments, refers to the container and its location within the segment's address space. The use of the term stripe refers to the same collection of shards as a segment, including how the shards are distributed along with the redundancy or parity information according to some embodiments.

[0117] A series of address space translations occurs throughout the storage system. At the top are directory entries (file names) that link to inodes. Inodes point into the media address space where data is logically stored. Media addresses may be mapped through a series of indirection media to spread the load of large files or implement data services such as deduplication or snapshots. Media addresses may be mapped through a series of indirection media to spread the load of large files or implement data services such as deduplication or snapshots. Segment addresses are then translated into physical flash locations. Physical flash locations have address ranges limited by the amount of flash in the system, according to some embodiments. Media addresses and segment addresses are logical containers, and in some embodiments use identifiers of 128 bits or more to be effectively infinite, with a likelihood of reuse calculated to be longer than the expected life of the system. Addresses from the logical containers are assigned hierarchically, in some embodiments. Initially, each non-volatile solid-state storage unit 152 may be assigned a range of address space. Within this allocated range, non-volatile solid-state storage 152 can allocate addresses without synchronization with other non-volatile solid-state storage 152.

[0118] Data and metadata are stored per storage device and in a collection of basic storage layouts optimized for varying workload patterns. These layouts incorporate multiple redundancy schemes, compression formats, and indexing algorithms. Some of these layouts store information about authorities and authority masters, while others store file metadata and file data. Redundancy schemes include error correction codes to tolerate corrupted bits within a single storage device (e.g., a NAND flash chip), erasure codes to tolerate failures of multiple storage nodes, and replication schemes to tolerate data center or regional failures. In some embodiments, low-density parity check (LDPC) codes are used within a single storage unit. Reed-Solomon coding is used in storage clusters, and mirroring is used in storage grids in some embodiments. Metadata may be stored using an ordered log-structured index (e.g., a log-structured merge tree), and large data may not be stored in a log-structured layout.

[0119] To maintain consistency across multiple copies of an entity, storage nodes implicitly agree through computation on two things: (1) the authority that contains the entity, and (2) the storage node that contains the authority. The assignment of entities to authorities may be done by pseudo-randomly assigning entities to authorities, by dividing entities into ranges based on an externally generated key, or by placing a single entity into each authority. Examples of pseudo-random schemes are linear hashing and the Replication Under Scalable Hashing ("RUSH") family of hashes, which includes Controlled Replication Under Scalable Hashing ("CRUSH"). In some embodiments, pseudo-random assignment is utilized only to assign authorities to nodes, since the set of nodes may change. The set of authorities cannot change, and therefore any subjective function may be applied in these embodiments. Some placement schemes automatically place authorities on storage nodes. While other placement schemes rely on explicit mapping of authorities to storage nodes, in some embodiments, a pseudo-random scheme is utilized to map each authority to a set of candidate authority owners. A pseudo-random data distribution function associated with CRUSH may assign authorities to storage nodes and create a list of locations where authorities are assigned. Each storage node may have a copy of the pseudo-random data distribution function and arrive at the same calculation for distributing and later discovering or locating authorities. Each pseudo-random scheme requires a reachable set of storage nodes as input in some embodiments to infer the same target node. Once an entity is placed into an authority, it may be stored on a physical device so that expected failures do not result in unexpected data loss. In some embodiments, a rebalancing algorithm attempts to store copies of all entities within the authority in the same layout and on the same set of machines.

[0120] Examples of anticipated failures include device failure, machine theft, data center fire, and regional disasters such as nuclear or geological events. Various failures result in different levels of acceptable data loss. In some embodiments, theft of a storage node does not affect the security or reliability of the system. On the other hand, depending on the system configuration, a regional event could result in no data loss, loss of a few seconds or minutes of updates, or even complete data loss.

[0121] In embodiments, data placement for storage redundancy is independent of authority placement for data consistency. In some embodiments, storage nodes containing authorities do not contain any persistent storage devices. Instead, the storage nodes are connected to non-volatile solid-state storage units that do not contain authorities. The communication interconnect between the storage nodes and the non-volatile solid-state storage units is comprised of multiple communication technologies and has uneven performance and fault tolerance characteristics. In some embodiments, as described above, the non-volatile solid-state storage units are connected to the storage nodes via PCI Express, and the storage nodes are connected to each other within a single chassis using an Ethernet backplane, and the chassis are connected to each other to form a storage cluster. The storage cluster, in some embodiments, is connected to clients using Ethernet or Fibre Channel. When multiple storage clusters are configured in a storage grid, the multiple storage clusters are connected using the Internet or other long-distance networking links, such as "macro-scale" links or private links that do not traverse the Internet.

[0122] The authority owner has the exclusive right to modify entities, move entities from one non-volatile solid-state storage unit to another, and add and delete copies of entities. This allows basic data redundancy to be maintained. When an authority owner fails, is decommissioned, or becomes overloaded, the authority is transferred to a new storage node. Transient failures make it important to ensure that all non-faulty machines agree on the new authority location. Ambiguity caused by transient failures can be resolved manually by a remote system administrator or by a local hardware administrator (e.g., by physically removing the failed machine from the cluster or pressing a button on the failed machine), or automatically by a consensus protocol such as Paxos or a hot-warm failover scheme. In some embodiments, a consensus protocol is used and failover is automatic. If too many failures or replication events occur in too short a period of time, the system enters a self-preservation mode, suspending replication and data movement activity until an administrator intervenes according to some embodiments.

[0123] As authorities are transferred between storage nodes and authority owners update their authority entities, the system transfers messages between storage nodes and non-volatile solid-state storage units. Regarding persistent messages, messages with different purposes are of different types. Depending on the message type, the system maintains different ordering and durability guarantees. As persistent messages are processed, they are temporarily stored in multiple persistent and non-persistent storage hardware technologies. In some embodiments, messages are stored in RAM, NVRAM, and NAND flash devices, and various protocols are used to efficiently use each storage medium. Latency-sensitive client requests may be persisted in replicated NVRAM and then later in the NAND, while background rebalancing operations are persisted directly to the NAND.

[0124] Persistent messages are persistently stored before being sent. This allows the system to continue servicing client requests despite failures and component replacement. Many hardware components contain unique identifiers that are visible to system administrators, manufacturers, the hardware supply chain, and ongoing monitoring and quality control infrastructure, but applications running on top of the infrastructure addresses virtualize the addresses. These virtualized addresses do not change over the life of the storage system despite component failures and replacements. This allows components of the storage system to be replaced over time without reconfiguration or interruption to the processing of client requests. In other words, the system supports non-disruptive updates.

[0125] In some embodiments, virtualized addresses are stored with full redundancy. A continuous monitoring system correlates hardware and software status and hardware identifiers. This allows for the detection and prediction of failures due to defective components and manufacturing details. The monitoring system also, in some embodiments, allows for the proactive transfer of authorities and entities away from affected devices before a failure occurs by removing components from critical paths.

2C is a multi-level block diagram illustrating the contents of a storage node 150 and the contents of the non-volatile solid-state storage 152 of the storage node 150. Data is communicated to and from the storage node 150 by a network interface controller ("NIC") 202, in some embodiments. Each storage node 150 includes a CPU 156 and one or more non-volatile solid-state storage devices 152, as described above. Moving down one level in FIG. 2C, each non-volatile solid-state storage device 152 includes a non-volatile random access memory ("NVRAM") 204 and a relatively fast non-volatile solid-state memory such as flash memory 206. In some embodiments, the NVRAM 204 may be a component that does not require program/erase cycles (DRAM, MRAM, PCM), and may be memory that can support being written to much more frequently than the memory is read. 2C , NVRAM 204 is implemented in one embodiment as a high-speed volatile memory, such as dynamic random access memory (DRAM) 216, backed up by energy storage 218. Energy storage 218 provides sufficient power to keep DRAM 216 powered long enough for the contents to be transferred to flash memory 206 in the event of a power outage. In some embodiments, energy storage 218 is a capacitor, supercapacitor, electron, or other device that provides an adequate supply of energy sufficient to enable the transfer of the contents of DRAM 216 to a stable storage medium in the event of a power loss. Flash memory 206 is implemented as multiple flash dies 222, which may also be referred to as a package of flash dies 222 or an array of flash dies 222. It should be understood that flash dies 222 could be packaged in any of several ways, such as a single die per package, multiple dies per package (i.e., a multi-chip package), in a hybrid package, as bare dies on a printed circuit board or other substrate, as encapsulated dies, etc. In the illustrated embodiment, non-volatile solid-state storage 152 has a controller 212 or other processor and input/output (I/O) ports 210 coupled to controller 212. I/O ports 210 are coupled to CPU 156 and/or network interface controller 202 of flash storage node 150. Flash input/output (I/O) ports 220 are coupled to flash die 222, and direct memory access unit (DMA) 214 is coupled to controller 212, DRAM 216, and flash die 222. In the illustrated embodiment, I/O ports 210, controller 212, DMA unit 214, and flash I/O ports 220 are implemented on a programmable logic device ("PLD") 208, such as, for example, a field programmable gate array (FPGA). In this embodiment, each flash die 222 has pages organized as 16 kB (kilobyte) pages 224 and registers 226 through which data can be written to or read from the flash die 222. In additional embodiments, other types of solid-state memory are used instead of or in addition to the flash memory shown in the flash die 222.

[0127] In various embodiments disclosed herein, a storage cluster 161 may generally be contrasted with a storage array. Storage nodes 150 are part of a collection that makes up the storage cluster 161. Each storage node 150 owns a slice of data and the computing required to provide the data. Multiple storage nodes 150 cooperate to store and retrieve data. Storage memory or storage devices typically used in a storage array are not primarily involved in processing or manipulating data. The storage memory or storage devices in a storage array receive commands to read, write, or erase data. The storage memory or storage devices in a storage array are unaware of the larger system in which they are embedded or what the data represents. The storage memory or storage devices in a storage array may include various types of storage memory, such as RAM, solid-state drives, hard disk drives, etc. The storage units 152 described herein have multiple interfaces that are simultaneously active and serve multiple purposes. In some embodiments, some of the functionality of storage node 150 is shifted into storage unit 152, transforming storage unit 152 into a combination of storage unit 152 and storage node 150. Putting the computation (on the stored data) into storage unit 152 places this computation closer to the data itself. Various system embodiments have a hierarchy of storage node tiers with different functionality. In contrast, in a storage array, the controller owns and knows everything about all of the data it manages within its shelves or storage devices. In storage cluster 161, multiple controllers of multiple storage units 152 and/or storage nodes 150 cooperate in various ways (e.g., for erasure coding, data sharding, metadata communication and redundancy, storage capacity expansion or contraction, data recovery, etc.) as described herein.

[0128] Figure 2D shows a storage server environment using an embodiment of the storage node 150 and storage unit 152 of Figures 2A-2C. In this version, each storage unit 152 has a processor such as a controller 212 (see Figure 2C), an FPGA (Field Programmable Gate Array), flash memory 206, and NVRAM 204 (which is DRAM 216 backed up by supercapacitors, see Figures 2B and 2C) on a PCIe (Peripheral Component Interconnect Express) board within the chassis 138 (see Figure 2A). The storage unit 152 may be implemented as a single board containing storage and may be the largest tolerable failure domain within the chassis. In some embodiments, up to two storage units 152 can fail and the device continues without data loss.

[0129] In some embodiments, physical storage is divided into regions named based on application usage. NVRAM 204 is a contiguous block of reserved memory within storage unit 152 DRAM 216 and backed by NAND flash. NVRAM 204 is logically divided into multiple memory regions that are written to for two as spools (e.g., spool_region). Space within the NVRAM 204 spools is managed independently by each authority 168. Each device provides an amount of storage space to each authority 168, which further manages the lifetime and allocation within that space. Examples of spools include distributed transactions or concepts. When primary power to storage unit 152 fails, an on-board supercapacitor provides short-term power holdup. During this holdup interval, the contents of NVRAM 204 are flushed to flash memory 206. Upon next power-on, the contents of NVRAM 204 are retrieved from flash memory 206.

[0130] With respect to storage unit controllers, logical "controller" responsibilities are distributed across each of the blades containing an authority 168. This distribution of logical control is shown in FIG. 2D as host controller 242, mid-tier controller 244, and storage unit controller(s) 246. While the parts may be physically located on the same blade, management of the control plane and storage plane is handled independently. Each authority 168 effectively functions as an independent controller. Each authority 168 provides its own data and metadata structures, its own background workers, and maintains its own lifecycle.

[0131] Figure 2E is a blade 252 hardware block diagram showing the control plane 254, compute plane, and storage planes 256, 258, as well as the authority 168 that interacts with the underlying physical resources using an embodiment of the storage node 150 and storage unit 152 of Figures 2A-2C in the storage server environment of Figure 2D. The control plane 254 is partitioned into several authorities 168 that can use the computational resources in the compute plane 256 to run on any of the blades 252. The storage plane 258 is partitioned into a collection of devices, each of which provides access to flash 206 and NVRAM 204 resources.

2E, authorities 168 interact with underlying physical resources (i.e., devices). From the perspective of the authority 168, its resources are striped across all of the physical devices. From the perspective of the device, it provides resources to all authorities 168, regardless of where the authority happens to be running. Each authority 168 has assigned or is assigned one or more partitions 260 of storage memory within the storage unit 152, such as partitions 260 of flash memory 206 and NVRAM 204. Each authority 168 uses its assigned partitions 260 to write or read user data. Authorities may be associated with different amounts of physical storage in the system. For example, one authority 168 may have more partitions 260 or larger-sized partitions 260 on one or more storage units 152 than one or more other authorities 168.

2F illustrates elasticity software layers for a blade 252 of a storage cluster, according to some embodiments. In an elasticity architecture, the elasticity software is symmetric. That is, each blade's compute module 270 executes the three identical layers of the process illustrated in FIG. 2F. The storage manager 274 executes read and write requests from other blades 252 to data and metadata stored in the local storage unit 152, NVRAM 204, and flash 206. The authority 168 fulfills client requests by issuing the necessary reads and writes to the blade 252 on whose storage unit 152 the corresponding data or metadata resides. The endpoint 272 analyzes client connection requests received from the switch fabric 146 oversight software, relays the client connection request to the authority 168 responsible for fulfillment, and relays the authority 168's response to the client. The symmetrical three-tier architecture enables a high degree of concurrency in the storage system. Elasticity scales out efficiently and reliably in these embodiments. Furthermore, Elasticity implements a unique scale-out strategy that maximizes concurrency by balancing work evenly across all resources regardless of client access patterns, eliminating much of the need for cross-blade coordination that typically occurs with traditional distributed locks.

[0134] Still referring to FIG. 2F, the authority 168, executing on the compute module 270 of the blade 252, performs the internal operations required to service client requests. One feature of elasticity is that the authority 168 is stateless; that is, the authority caches active data and metadata in its own blade 252's DRAM for fast access, but the authority stores any updates to its NVRAM 204 partitions on three separate blades 252 until the updates are written to flash 206. All storage system writes to NVRAM 204 are, in some embodiments, triplicate for partitions on three separate blades 252. With triple-mirrored NVRAM 204 and persistent storage protected by parity and Reed-Solomon RAID checksums, the storage system can survive the simultaneous failure of two blades 252 without losing data, metadata, or access to either.

[0135] Because authorities 168 are stateless, authorities can migrate between blades 252. Each authority 168 has a unique identifier. NVRAM 204 and flash 206 partitions are associated with the identifier of the authority 168, not with the blade 252 on which some authority is running. Thus, when an authority 168 migrates, the authority 168 continues to manage the same storage partitions from its new location. When a new blade 252 is installed in a storage cluster embodiment, the system partitions the new blade's 252's storage for use by the system's authorities 168, migrates selected authorities 168 to the new blade 252, and automatically rebalances the load by starting endpoints 272 on the new blade 252 and including them in the switch fabric 146's client connection distribution algorithm.

[0136] From its new location, the migrated authority 168 persists the contents of its NVRAM 204 partition on flash 206, processes read and write requests from other authorities 168, and serves client requests directed to it by endpoints 272. Similarly, if a blade 252 fails or is removed, the system redistributes its authorities 168 among the remaining blades 252 in the system. The redistributed authorities 168 continue to perform their original functions from their new locations.

2G illustrates the storage resources of an authority 168 and a blade 252 of a storage cluster, according to some embodiments. Each authority 168 is exclusively responsible for a partition of flash 206 and NVRAM 204 on each blade 252. The authority 168 manages the contents and integrity of its partition independently of other authorities 168. The authority 168 compresses incoming data, temporarily stores it in its NVRAM 204 partition, and then hardens, RAID-protects, and persists the data in a segment of storage within its flash 206 partition. As the authority 168 writes data to the flash 206, the storage manager 274 performs the necessary flash transformations to optimize write performance and maximize media lifespan. In the background, the authority 168 "garbage collects," or reclaims, the space occupied by data that clients make obsolete by overwriting the data. It should be understood that because the partitions of authority 168 have no common elements, there is no need for distributed locking between clients and writes or to perform background functions.

[0138] The embodiments described herein may utilize a variety of software protocols, communication protocols, and/or networking protocols. Additionally, hardware and/or software configurations may be adjusted to accommodate the various protocols. For example, embodiments may utilize Active Directory, a database-based system that provides authentication, directory, policy, and other services in a WINDOWS™ environment. In these embodiments, LDAP (Lightweight Directory Access Protocol) is an example application protocol for querying and modifying items in a directory service provider such as Active Directory. In some embodiments, a Network Lock Manager ("NLM") is utilized as a mechanism that provides System V-style advisory files and works in conjunction with the Network File System ("NFS") to record locks over the network. The Server Message Block ("SMB") protocol, one version of which is also known as the Common Internet File System ("CIFS"), may be integrated with the storage systems described herein. SMP operates as an application-layer network protocol typically used to provide shared access to files, printers, and serial ports, as well as various communications between nodes on a network. SMB also provides an authenticated inter-process communication mechanism. AMAZON™ S3 (Simple Storage Service) is a web service provided by Amazon Web Services, and the system described herein may interface with Amazon S3 through web service interfaces (REST (Representable State Transfer), SOAP (Simple Object Access Protocol), and BitTorrent). A RESTful API (Application Programming Interface) breaks down transactions to create a series of small modules, each addressing a specific fundamental part of the transaction. The control or permissions provided in these embodiments, particularly for object data, may include the use of access control lists ("ACLs"). An ACL is a list of permissions granted to an object; an ACL specifies not only which operations are allowed on a given object, but also which user or system processes are allowed to access the object. The system provides an identification and location system for computers on the network and utilizes Internet Protocol version 6 ("IPV6"), as well as IPv4, as the communications protocol for sending traffic across the Internet. Routing of packets between networked systems may include equal-cost multipath routing ("ECMP"). ECMP is a routing strategy in which next-hop packets forwarded to a single destination may occur on multiple "best routes" that rank for top place in a routing metric calculation. Because multipath routing is a hop-by-hop decision limited to a single router, multipath routing can be used in conjunction with most routing protocols. The software may support multitenancy, an architecture in which a single instance of a software application serves multiple customers. Each customer may be referred to as a tenant. Tenants may be given the ability to customize some parts of the application, but in some embodiments, may not customize the application's code. Embodiments may maintain an audit log. An audit log is a document that records events in a computing system. In addition to documenting which resources were accessed, audit log entries typically include destination and source addresses, timestamps, and user login information for compliance with various regulations. Embodiments may support various key management policies, such as encryption key rotation. Additionally, the system may support a dynamic root password or some variant of a dynamically changing password.

3A illustrates a diagram of a storage system 306 coupled for data communication with a cloud service provider 302 in accordance with some embodiments of the present disclosure. While not shown in greater detail, the storage system 306 shown in FIG. 3A may be similar to the storage systems described above with respect to FIGS. 1A-1D and 2A-2G. In some embodiments, the storage system 306 shown in FIG. 3A may be implemented as a storage system including unbalanced active/active controllers, a storage system including balanced active/active controllers, a storage system including active/active controllers in which less than all of the resources of each controller are utilized such that each controller has spare resources that can be used to support failover, a storage system including fully active/active controllers, a storage system including data set separation controllers, a storage system including a dual-tier architecture with a front-end controller and a back-end unified storage controller, a storage system including a scale-out cluster of dual-controller arrays, and combinations of such embodiments.

3A, storage system 306 is coupled to cloud service provider 302 via data communications link 304. Data communications link 304 may be implemented as a dedicated data communications link, as a data communications path provided through the use of one of a data communications network such as a wide area network ("WAN") or a local area network ("LAN"), or some other mechanism capable of transporting digital information between storage system 306 and cloud service provider 302. Such data communications link 304 may be entirely wired, entirely wireless, or some aggregation of wired and wireless data communications paths. In such an example, digital information may be exchanged between storage system 306 and cloud service provider 302 via data communications link 304 using one or more data communications protocols. For example, digital information may be exchanged between storage system 306 and cloud service provider 302 over data communications link 304 using Handheld Device Transfer Protocol ("HDTP"), Hypertext Transfer Protocol ("HTTP"), Internet Protocol ("IP"), Real-Time Transport Protocol ("RTP"), Transmission Control Protocol ("TCP"), User Datagram Protocol ("UDP"), Wireless Application Protocol ("WAP"), or other protocols.

[0141] The cloud service provider 302 shown in FIG. 3A may be implemented as a system and computing environment that provides services to users of the cloud service provider 302 through the sharing of computing resources, for example, via a data communication link 304. The cloud service provider 302 may provide on-demand access to a shared pool of configurable computing resources, such as computer networks, servers, storage, applications, and services. The shared pool of configurable resources may be quickly set up and released to users of the cloud service provider 302 with minimal administrative effort. Generally, users of the cloud service provider 302 are unaware of the exact computing resources utilized by the cloud service provider 302 to provide services. While in many cases, such cloud service providers 302 may be accessible via the Internet, readers skilled in the art will recognize that any system that conceptualizes the use of shared resources to provide services to users over any data communication link may be considered a cloud service provider 302.

3A, cloud service provider 302 may be configured to provide various services to storage system 306 and users of storage system 306 through various service model implementations. For example, cloud service provider 302 may be configured to provide services to storage system 306 and users of storage system 306 through an infrastructure-as-a-service ("IaaS") service model implementation in which cloud service provider 302 provides computing infrastructure, such as virtual machines and other resources, as a service to subscribers. Furthermore, cloud service provider 302 may be configured to provide services to storage system 306 and users of storage system 306 through a platform-as-a-service ("PaaS") service model implementation in which cloud service provider 302 provides a development environment to application developers. Such a development environment may include, for example, an operating system, a programming language execution environment, a database, a web server, or other components that may be used by application developers to develop and run software solutions on a cloud platform. Additionally, cloud service provider 302 may be configured to provide services to storage system 306 and users of storage system 306 through an implementation of a software-as-a-service ("SaaS") service model in which cloud service provider 302 provides application software, databases, as well as a platform used to run applications to storage system 306 and users of storage system 306, providing on-demand software to storage system 306 and users of storage system 306, eliminating the need to install and run applications on local computers and simplifying application maintenance and support. Cloud service provider 302 may further be configured to provide services to storage system 306 and users of storage system 306 through an implementation of an authentication-as-a-service ("AaaS") service model in which cloud service provider 302 provides authentication services that can be used to secure access to applications, data sources, or other resources. Cloud service provider 302 may also be configured to provide services to storage system 306 and users of storage system 306 through an implementation of a storage-as-a-service model in which cloud service provider 302 provides access to its storage infrastructure for use by storage system 306 and users of storage system 306. The reader will understand that cloud service provider 302 may be configured to provide additional services to storage system 306 and users of storage system 306 through the implementation of additional service models, as the service models described above are included for illustrative purposes only and do not represent any limitations on the services that may be provided by cloud service provider 302 or on the service models that may be implemented by cloud service provider 302.

3A, cloud service provider 302 may be implemented, for example, as a private cloud, a public cloud, or a combination of private and public clouds. In embodiments in which cloud service provider 302 is implemented as a private cloud, cloud service provider 302 may focus on providing services to a single organization rather than serving multiple organizations. In embodiments in which cloud service provider 302 is implemented as a public cloud, cloud service provider 302 may provide services to multiple organizations. Public and private cloud deployment models may differ and may involve various advantages and disadvantages. For example, because public cloud deployments require the sharing of computing infrastructure across different organizations, such deployments may not be ideal for organizations with security concerns, mission-critical workloads, uptime requirements, etc. While private cloud deployments can address some of these issues, private cloud deployments may require on-premise staff to manage the private cloud. In yet another alternative embodiment, cloud service provider 302 may be implemented as a hybrid cloud deployment of private and public cloud services.

3A , the reader will understand that additional hardware and software components may be necessary to facilitate delivery of cloud services to the storage system 306 and users of the storage system 306. For example, the storage system 306 may be coupled to (or include) a cloud storage gateway. Such a cloud storage gateway may be implemented, for example, as a hardware- or software-based appliance located at the facility containing the storage system 306. Such a cloud storage gateway may act as a bridge between local applications running on the storage array 306 and remote, cloud-based storage leveraged by the storage array 306. Using a cloud storage gateway, an organization can move temporary iSCSI or NAS data to the cloud service provider 302, allowing the organization to conserve space in its on-premises storage system. Such a cloud storage gateway may be configured to emulate a disk array, block-based device, file server, or other storage system that can translate SCSI commands, file server commands, or other appropriate commands into a RESTful storage protocol that facilitates communication with the cloud service provider 302.

[0145] A cloud migration process may occur during which data, applications, or other elements are moved from the organization's local system (or from another cloud environment) to cloud service provider 302 to enable storage system 306 and users of storage system 306 to take advantage of services offered by cloud service provider 302. To successfully migrate the data, applications, or other elements to the cloud service provider's 302 environment, middleware such as a cloud migration tool may be utilized to bridge the gap between the cloud service provider's 302 environment and the organization's environment. Such cloud migration tools may also be configured to address security concerns associated with transmitting sensitive data over a data communications network to cloud service provider 302, as well as potentially high network costs and long migration times associated with migrating large amounts of data to cloud service provider 302. To further enable storage system 306 and users of storage system 306 to take advantage of services offered by cloud service provider 302, a cloud orchestrator may be used to prepare and coordinate automated tasks to create an integrated process or workflow. Such a cloud orchestrator may perform tasks such as configuring various components, whether those components are cloud or on-premise components, as well as managing the interconnections between such components. The cloud orchestrator can simplify communication and connections between components to ensure that links are properly configured and maintained.

3A, and as briefly described above, cloud service provider 302 may be configured to provide services to storage system 306 and users of storage system 306 through the use of a SaaS service model in which cloud service provider 302 provides application software, databases, as well as the platform used to run the applications to storage system 306 and users of storage system 306, providing on-demand software to storage system 306 and users of storage system 306, eliminating the need to install and run applications on local computers, which may simplify application maintenance and support. Such applications may take many forms in accordance with various embodiments of the present disclosure. For example, cloud service provider 302 may be configured to provide storage system 306 and users of storage system 306 with access to a data analysis application. Such a data analysis application may be configured, for example, to receive telemetry data communicated by storage system 306 to its servers. Such telemetry data may describe various operating characteristics of storage system 306 and may be analyzed, for example, to determine the health of storage system 306, to identify the workload running on storage system 306, to predict when storage system 306 will run out of various resources, and to recommend configuration changes, hardware or software upgrades, workflow transitions, or other actions that may improve the operation of storage system 306.

[0147] Cloud service provider 302 may also be configured to provide storage system 306 and users of storage system 306 with access to virtualized computing environments. Such virtualized computing environments may be implemented, for example, as virtual machines or other virtualized computer hardware platforms, virtual storage devices, virtualized computer network resources, etc. Examples of such virtualized environments may include virtual machines created to emulate actual computers, virtualized desktop environments that separate logical desktops from physical machines, virtualized file systems that enable uniform access to different types of concrete file systems, and many others.

[0148] For further explanation, FIG. 3B illustrates a diagram of a storage system 306 according to some embodiments of the present disclosure. Although not described in great detail, the storage system 306 shown in FIG. 3B may be similar to the storage systems described above with reference to FIGS. 1A-1D and 2A-2G, as the storage system may include many of the components described above.

3B may include storage resources 308, which may be implemented in many ways. For example, in some embodiments, storage resources 308 may include nanoRAM or another form of nonvolatile random access memory utilizing carbon nanotubes deposited on a substrate. In some embodiments, storage resources 308 may include 3D cross-point nonvolatile memory in which bit storage is based on bulk resistance changes, along with stackable cross-gridded data access arrays. In some embodiments, storage resources 308 may include flash memory, including single-level cell ("SLC") NAND flash, multi-level cell ("MLC") NAND flash, triple-level cell ("TLC") NAND flash, quad-level cell ("QLC") NAND flash, and others. In some embodiments, storage resources 308 may include nonvolatile magnetoresistive random access memory ("MRAM"), including spin-transfer torque ("STT") MRAM, in which data is stored using magnetic storage elements. In some embodiments, example storage resource 308 may include non-volatile phase change memory (“PCM”), which may have the ability to hold multiple bits in a single cell because the cell can achieve several distinct intermediate states. In some embodiments, storage resource 308 may include quantum memory, which allows for the storage and retrieval of optical quantum information. In some embodiments, example storage resource 308 may include resistive random access memory (“ReRAM”), in which data is stored by changing the resistance across a dielectric solid-state material. In some embodiments, storage resource 308 may include storage class memory (“SCM”), in which lid-state non-volatile memory may be fabricated at high density using some combination of sublithographic patterning techniques, multiple bits per cell, multiple layers of devices, etc. The reader will understand that other forms of computer memory and storage devices may be utilized by the above-described storage system, including DRAM, SRAM, EEPROM, universal memory, and many others. The storage resource 308 shown in FIG. 3A may be implemented in a variety of form factors, including, but not limited to, dual in-line memory modules ("DIMMs"), non-volatile dual in-line memory modules ("NVDIMMs"), M.2, U.2, and others.

[0150] The example storage system 306 shown in FIG. 3B may implement a variety of storage architectures. For example, storage systems according to some embodiments of the present disclosure may utilize block storage, where data is stored in blocks, with each block essentially performing the function of an individual hard drive. Storage systems according to some embodiments of the present disclosure may utilize object storage, where data is managed as objects. Each object may include the data itself, a variable amount of metadata, and a globally unique identifier, and object storage can be implemented at multiple levels (e.g., device level, system level, interface level). Storage systems according to some embodiments of the present disclosure utilize file storage, where data is stored in a hierarchical structure. Such data is stored in files and folders and is presented in the same format to both the system that stores it and the system that retrieves it.

[0151] The example storage system 306 shown in FIG. 3B may be implemented as a storage system in which additional storage resources can be added using a scale-up model, additional storage resources can be added using a scale-out model, or some combination thereof. In a scale-up model, additional storage may be added by adding additional storage devices. However, in a scale-out model, additional storage nodes may be added to a cluster of storage nodes, and such storage nodes may include additional processing resources, additional networking resources, etc.

3B also includes communications resources 310 that may be useful in facilitating data communication between components within the storage system 306, as well as between the storage system 306 and computing devices external to the storage system 306. The communications resources 310 may be configured to utilize a variety of different protocols and data communications fabrics to facilitate data communication between components within the storage system, as well as between computing devices external to the storage system. For example, the communications resources 310 may include Fibre Channel ("FC") technology, such as an FC fabric and FC protocol that may transport SCSI commands over an FC network. The communications resources 310 may also include FC over Ethernet ("FCoE") technology, in which FC frames are encapsulated and transmitted over an Ethernet network. The communications resources 310 may also include InfiniBand ("IB") technology, in which switched fabric technology is utilized to facilitate transmission between channel adapters. The communications resources 310 may also include NVM Express ("NVMe") and NVMe over Fabric ("NVMeoF") technologies, through which non-volatile storage media attached via a PCI Express ("PCIe") bus may be accessed. The communications resources 310 may also include mechanisms for accessing the storage resources 308 within the storage system 306 using Serial Attached SCSI ("SAS"), a Serial ATA ("SATA") bus interface for connecting the storage resources 308 within the storage system 306 to a host bus adapter within the storage system 306, and Internet Small Computer System Interface ("iScSI") technologies to provide block-level access to the storage resources 308 within the storage system 306 and other communications resources that may be useful for facilitating data communication between components within the storage system 306 as well as between the storage system 306 and computing devices external to the storage system 306.

[0153] The storage system 306 shown in FIG. 3B also includes processing resources 312 that may be useful in executing computer program instructions and performing other computational tasks within the storage system 306. The processing resources 312 may include one or more central processing units ("CPUs") as well as one or more application-specific integrated circuits ("ASICs") customized for some particular purpose. The processing resources 312 may also include one or more digital signal processors ("DSPs"), one or more field programmable gate arrays ("FPGAs"), one or more systems-on-chips ("SoCs"), or other forms of processing resources 312. The storage system 306 may utilize the storage resources 312 to perform various tasks, including, but not limited to, supporting the execution of software resources 314, which are described in more detail below.

3B also includes software resources 314 that, when executed by processing resources 312 within storage system 306, may perform various tasks. Software resources 314 may include, for example, one or more modules of computer program instructions that, when executed by processing resources 312 within storage system 306, are useful in implementing various data protection techniques to preserve the integrity of data stored within the storage system. The reader will understand that such data protection techniques may be implemented, for example, by system software running on computer hardware within the storage system, by a cloud service provider, or in other manners. Such data protection techniques may include, for example, data archiving techniques, in which data that is no longer in active use is moved to a separate storage device or separate storage system for long-term retention; data backup techniques, in which data stored in a storage system may be copied and stored elsewhere to avoid data loss in the event of equipment failure or some other form of catastrophic event with the storage system; data replication techniques, in which data stored in a storage system is replicated to another storage system so that the data is accessible via multiple storage systems; data snapshot techniques, in which the state of data in a storage system is captured at various points in time; data and database cloning techniques, through which copies of data and database replicas may be created, and other data protection techniques. Because a failure of a storage system may not result in the loss of data stored in the storage system, business continuity and disaster recovery objectives may be met through the use of such data protection techniques.

[0155] Software resources 314 may also include software useful in implementing software-defined storage ("SDS"). In such examples, software resources 314 may include one or more modules of computer program instructions that, when executed, are useful in policy-based provisioning and management of data storage independent of the underlying hardware. Such software resources 314 may be useful in implementing storage visualization to separate storage hardware from the software that manages the storage hardware.

[0156] Software resources 314 may also include software useful in facilitating and optimizing I/O operations directed to storage resources 308 within storage system 306. For example, software resources 314 may include software modules that implement various data reduction techniques, such as data compression, data deduplication, and others. Software resources 314 may also include software modules that intelligently group I/O operations together to facilitate better use of underlying storage resources 308, software modules that perform data migration operations to migrate data out of the storage system, and software modules that perform other operations. Such software resources 314 may be implemented as one or more software containers or in many other ways.

[0157] The reader will understand that the various components shown in FIG. 3B may be grouped into one or more optimized computing packages as a centralized infrastructure. Such a centralized infrastructure may include a pool of computing, storage, and networking resources that may be shared by multiple applications and collectively managed using policy-driven processes. Such a centralized infrastructure may minimize compatibility issues between the various components in storage system 306 while also reducing various costs associated with establishing and operating storage system 306. Such a centralized infrastructure may be implemented in a centralized infrastructure reference architecture, in standalone appliances, in a software-driven hyper-centralized approach (e.g., a hyper-centralized infrastructure), or in other ways.

[0158] The reader will appreciate that storage system 306 shown in FIG. 3B may be useful for supporting various types of software applications. For example, storage system 306 may be useful in supporting artificial intelligence ("AI") applications, database applications, DevOps projects, electronic design automation tools, event-driven software applications, high-performance computing applications, simulation applications, high-speed data capture and analysis applications, machine learning applications, media generation applications, media provisioning applications, picture archiving and communication systems ("PACS") applications, software development applications, virtual reality applications, augmented reality applications, and many other types of applications by providing storage resources for such applications.

[0159] The storage system described above may operate to support a wide variety of applications. Given the fact that the storage system includes computational resources, storage resources, and a wide variety of other resources, the storage system may be well suited to support resource-intensive applications, such as AI applications. Such AI applications may enable a device to recognize its environment and take actions that maximize its probability of success toward some goal. Examples of such AI applications may include IBM Watson, Microsoft Oxford, Google DeepMind, Baidu Minwa, and others. The storage system described above may also be well suited to support other types of resource-intensive applications, such as machine learning applications. Machine learning applications may perform various types of data analysis to automate analytical model building. Machine learning applications use algorithms that iteratively learn from data, allowing computers to learn without being explicitly programmed.

[0160] In addition to the resources already described, the storage systems described above may also include a graphics processing unit ("GPU"), sometimes referred to as a visual processing unit ("VPU"). Such a GPU may be implemented as specialized electronic circuitry that rapidly manipulates and modifies memory to accelerate the creation of images within a frame buffer intended for output to a display device. Such a GPU may be included within any of the computing devices that are part of the storage systems described above, including as one of many individually scalable components of the storage system; other examples of individually scalable components of such storage systems may include storage components, memory components, computational components (e.g., CPUs, FPGAs, ASICs), networking components, software components, and others. In addition to the GPU, the storage systems described above may also include a neural network processor ("NNP") for use in various aspects of neural network processing. Such NNPs may be used instead of (or in addition to) GPUs and may be independently scalable.

[0161] As mentioned above, the storage systems described herein may be configured to support artificial intelligence applications, machine learning applications, big data analytics applications, and many other types of applications. The rapid growth of these types of applications is driven by three technologies: deep learning (DL), GPU processors, and big data. Deep learning is a computing model that utilizes massively parallel neural networks inspired by the human brain. Instead of experts handcrafting software, deep learning models write their own software by learning from many examples. GPUs are modern processors with thousands of cores that are well suited to running algorithms that loosely represent the parallel nature of the human brain.

[0162] Advances in deep neural networks have ignited a new wave of algorithms and tools for data scientists to utilize that data with artificial intelligence (AI). With improved algorithms, larger datasets, and diverse frameworks (including open-source software libraries for machine learning across a range of tasks), data scientists are tackling new use cases such as autonomous vehicles, natural language processing, and many others. However, training deep neural networks requires both high-quality input data and large amounts of computation. GPUs are massively parallel processors that can operate on large amounts of data simultaneously. When coupled to multi-GPU clusters, high-throughput pipelines may be required to send input data from storage to the computation engine. Deep learning is more than simply building and training a model. There is also an entire data pipeline that must be designed for the scale, iteration, and experimentation required for data science teams to be successful.

[0163] Data is at the heart of modern AI and deep learning algorithms. Before training can begin, one problem that must be addressed revolves around collecting labeled data, which is critical for training accurate AI models. Full-scale AI deployments may be required to continuously collect, organize, transform, label, and store large amounts of data. Adding additional high-quality data points directly translates into more accurate models and better insights. Data samples may undergo a series of processing steps, including, but not limited to, 1) ingesting data from external sources into the training system and storing the data in a raw format; 2) organizing the data and converting it into a format convenient for training, including linking data samples to appropriate labels; 3) iteratively exploring parameters and models, rapidly testing them on smaller datasets, and focusing on the most promising models for pushing into the production cluster; 4) running a training phase to select random batches of input data containing both new and older samples and feeding the random batches to a production CPU server for calculations to update the model parameters; and 5) evaluating, including using a holdback portion of the data not used in training to evaluate model accuracy on holdout data. This lifecycle may apply for any type of parallelized machine learning, not just neural networks or deep learning. For example, a standard machine learning framework may rely on a CPU instead of a GPU, but the data ingestion and training workflow may be the same. Readers will understand that a single shared storage data hub creates a coordination point throughout the lifecycle without requiring extra data copies during the ingestion, preprocessing, and training stages. Ingested data is rarely used for just one purpose, and shared storage provides the flexibility to train multiple different models or apply traditional analytics to the data.

[0164] The reader understands that each stage of an AI data pipeline may have varying requirements from a data hub (e.g., a storage system or collection of storage systems). A scale-out storage system must deliver uncompromising performance for all types of access types and patterns—from small metadata to heavy, large files, random access patterns to sequential access patterns, and low concurrency to high concurrency. Because the system may serve unstructured workloads, the storage system described above may serve as an ideal AI data hub. In the first stage, data is ideally ingested and stored on the same data hub used by subsequent stages to avoid excessive data copying. The next two steps can be run on standard compute servers, optionally including GPUs, and then in the fourth and final stage, full training production jobs are run on powerful CPU-accelerated servers. Often, there is a production pipeline alongside an experimental pipeline running on the same dataset. Furthermore, GPU-accelerated servers can be used independently for different models or can be combined together to train on a single larger model, spanning multiple systems for distributed training. If the shared storage layer is slow, then data must be copied to local storage for each stage, resulting in wasted time staging data on different servers. An ideal data hub for an AI training pipeline would provide performance similar to data stored locally on the server nodes, while also possessing the simplicity and performance to allow all pipeline stages to operate simultaneously.

[0165] Data scientists work to improve the usability of trained models through a variety of methods: more data, better data, smarter training, and deeper models. Often, there are teams of data scientists sharing the same dataset and working in parallel to create new and improved training models. Often, there are teams of data scientists working through these stages simultaneously on the same shared dataset. Multiple concurrent workloads of data processing, experimentation, and full-scale training layer the demands of multiple access patterns on top of the storage layer. In other words, rather than being able to satisfy large file reads, storage must handle a mixture of large and small file reads and writes. Finally, with multiple data scientists exploring datasets and models, storing data in its native format can be crucial to provide each user with the flexibility to transform, organize, and use the data in their own way. The storage system described above can provide a natural shared storage home for datasets, data protection redundancy (e.g., by using RAID 6), and the performance necessary to be a common access point for multiple developers and multiple experiments. Using the storage system described above can avoid the need to carefully copy subsets of data for local work, saving both engineering and CPU-accelerated server time. These copies become an ever-growing burden as raw datasets persist and desired transformations are constantly updated and changing.

[0166] The reader will understand that the underlying reason why deep learning is experiencing such rapid success is the continuous improvement of models with larger dataset sizes. In contrast, older machine learning algorithms, such as logistic regression, stop improving accuracy at smaller dataset sizes. In this way, the separation of computational and storage resources also allows for independent scaling of each tier, avoiding much of the complexity inherent in managing both together. As dataset sizes grow or new datasets are considered, the scale-out storage system must be able to easily expand. Similarly, if more concurrent training is required, additional GPUs or other computational resources can be added without concern for their internal storage. Additionally, the storage system described above may make it easier to build, operate, and grow AI systems due to the random read bandwidth provided by the storage system, the ability of the storage system to randomly read small files (50 KB) at high speed (meaning that no extra effort is required to consolidate individual data points to create larger storage-friendly files), the ability of the storage system to scale capacity and performance as data sets grow or throughput requirements increase, the ability of the storage system to support files or objects, the ability of the storage system to tune performance for large or small files (i.e., the user does not need to set up a file system), the ability of the storage system to support non-destructive hardware and software updates even during production model training, and for many other reasons.

[0167] Many types of input, including text, audio, or images, are natively stored as small files, so small file performance at the storage layer can be important. If the storage layer does not handle small files well, extra steps are required to preprocess and group samples into larger files. Storage built on top of a spinning platter that relies on SSDs as a cache layer may fall short of the required performance. Because training with random input batches yields more accurate models, the entire dataset must be accessible at full performance. SSD caches only provide high performance for a small subset of data and are ineffective at hiding the latency of the spinning drives.

[0168] The reader will understand that the storage systems described above may be configured to support the storage of blockchains (among other types of data). Such blockchains may be implemented as a continuously growing list of records called blocks that are linked and secured using cryptography. Each block in a blockchain may contain a hash pointer as a link to the previous block, a timestamp, transaction data, etc. Blockchains may be designed to be resistant to data modification and can serve as an open distributed ledger that can record transactions between two parties efficiently, in a verifiable, and permanent manner. This makes blockchains potentially suitable for recording events, medical records, and other record-keeping activities such as identity management, transaction processing, and the like.

The reader will further understand that in some embodiments, the storage systems described above may be paired with other resources to support the applications described above. For example, one infrastructure might include primary computing in the form of servers and workstations specialized in using general-purpose computing on graphics processing units (GPGPUs) to accelerate deep learning applications, interconnected to a compute engine to train parameters for deep neural networks. Each system may have Ethernet external connectivity, InfiniBand external connectivity, some other form of external connectivity, or some combination thereof. In such an example, the GPUs might be grouped for a single large training run or used independently to train multiple models. The infrastructure might also include a storage system, such as those described above to provide a scale-out all-flash file or object store, with data accessible through high-performance protocols such as NFS, S3, etc. The infrastructure might also include redundant top-of-rack Ethernet switches connected to the storage and compute via ports in MLAG port channels for redundancy. The infrastructure may also include additional computation in the form of white-box servers, optionally using GPUs, for data ingestion, pre-processing, and model debugging. The reader understands that additional infrastructure is also possible.

[0170] The reader will understand that the above-described system may be more suitable for the above-described applications than other systems, which may include distributed direct-attached storage (DDAS) solutions deployed on server nodes. Such DDAS solutions may be built to handle larger, less sequential accesses, but may not be able to handle smaller, random accesses as well. The reader will understand that the above-described storage system may be utilized to provide the above-described applications with a platform that is preferable to leveraging cloud-based resources, or otherwise as part of a platform to support the above-described applications, because the storage system may be included in an on-site or in-house infrastructure that is more secure, more locally and internally managed, and more robust in feature set and performance. For example, services built on a platform such as IBM's Watson may require a company to distribute personal user information, such as financial transaction information or identifiable patient records, to other institutions. Thus, a cloud-based offering of AI as a service may be less desirable than AI that is managed internally and delivered as a service supported by a storage system, such as the storage system described above, for a variety of technical as well as business reasons.

[0171] The reader will understand that the storage system described above may be configured to support other AI-related tools, either alone or in conjunction with other computing machines. For example, the storage system may utilize tools such as ONXX or other open neural network exchange formats that facilitate the transfer of models created in different AI frameworks. Similarly, the storage system may be configured to support tools such as Amazon's Gluon, which allows developers to prototype, build, and train deep learning models.

[0172] The reader will further understand that the storage systems described above may be deployed as edge solutions. Such edge solutions may be implemented to optimize cloud computing systems by performing data processing at the edge of the network, closer to the source of the data. Edge computing may push applications, data, and computing power (i.e., services) from a centralized point to the logical edge of the network. Using an edge solution, such as the storage systems described above, computational tasks may be performed using the computational resources provided by such storage systems, data may be stored using the storage resources of the storage systems, and cloud-based services may be accessed using various resources of the storage systems (including networking resources). By performing computational tasks at edge solutions, storing data at edge solutions, and generally utilizing edge solutions, consumption of expensive cloud-based resources may be avoided, and in fact, performance improvements may be experienced with greater reliance on cloud-based resources.

[0173] While many tasks may benefit from leveraging edge solutions, some specific applications may be particularly suited for deployment in such environments. For example, devices such as drones, autonomous vehicles, robots, and others may require high-speed processing—so fast, in fact, that escalating data back to a cloud environment for data processing support may simply be too slow. Similarly, machines such as locomotives and gas turbines, which generate large amounts of information through the use of abundant data-generating sensors, may benefit from the high-speed data processing capabilities of edge solutions. As an additional example, some IoT devices, such as connected video cameras, may not be suitable for leveraging cloud-based resources because transmitting data to the cloud may be impractical (not just from a privacy, security, or financial perspective) simply due to the sheer volume of data involved. Thus, many tasks that rely on data processing, storage, or communication may be better suited by platforms that include edge solutions, such as the storage systems described above.

[0174] Consider the specific example of inventory management in a warehouse, distribution center, or similar location. A large inventory operation, freight storage operation, shipping operation, order fulfillment operation, manufacturing operation, or other operation has large amounts of inventory on the inventory shelves and high-resolution digital cameras that produce a large firehose of data. All of this data may be captured into an image processing system that can reduce the amount of data into a firehose of smaller data. All of the smaller data may be stored on-premise in storage. The on-premise storage at the edge of the facility may be coupled to the cloud for external reporting, real-time control, and cloud storage. Inventory management may be performed using the results of the image processing, whereby inventory is tracked on the shelves, replenished, moved, shipped, revised with new products, or discontinued/obsolete products are removed, etc. The above situation is a prime candidate for the configurable processing system and storage system embodiments described above. A combination of compute-specific blades and offload blades suitable for image processing, perhaps with deep learning on offload FPGAs or offload custom blade(s), could take the firehose of massive amounts of data from all of the digital cameras and create a firehose of smaller data. All of the smaller data would then be stored by the storage node, which would work with whatever combination of storage blade types best handles the data flow. This is an example of storage and functionality acceleration and consolidation. Depending on the need for external communication with and processing within the cloud, as well as the reliability of network connections and cloud resources, the system would be sized for managing bursty workloads and storage and computation with variable conductivity reliability. Also, depending on other inventory management aspects, the system could be configured for scheduling and resource management in a hybrid edge/cloud environment.

[0175] The storage systems described above may also be optimized for use in big data analytics. Big data analytics may be broadly described as the process of examining large and varied data sets to uncover hidden patterns, unknown correlations, market trends, customer preferences, and other useful information that can help organizations make more informed business decisions. Big data analytics applications enable data scientists, predictive modelers, statisticians, and other analytics professionals to analyze ever-increasing volumes of structured transactional data, along with other forms of data often left untapped by traditional business intelligence (BI) and analytics programs. As part of that process, semi-structured and unstructured data, such as internet clickstream data, web server logs, social media content, text from customer emails and survey responses, mobile phone call detail records, IoT sensor data, and other data, may be converted into a structured form. Big data analytics is a form of advanced analytics that requires complex applications involving elements such as predictive models, statistical algorithms, and what-if analysis powered by high-performance analytical systems.

[0176] The storage system described above may also support applications (including implementation as a system interface) that perform tasks in response to human speech. For example, the storage system may support running intelligent personal assistant applications such as Amazon's Alexa, Apple Siri, Google Voice, Samsung Bixby, Microsoft Cortana, and others. While the examples described above utilize speech as input, the storage system described above may also support chatbots, talkbots, chatterbots, or artificial conversational entities configured to conduct conversations via auditory or textual methods. Similarly, the storage system may actually run such applications to enable users, such as system administrators, to interact with the storage system via speech. In embodiments of the present disclosure, such applications may be utilized as an interface for a variety of system management operations, but generally, such applications may provide voice interaction, music playback, to-do list creation, alarm setting, podcast streaming, audiobook playback, and weather, traffic, and other real-time information such as news.

[0177] The storage systems described above may also implement AI platforms to achieve the vision of self-driving storage. Such AI platforms may be configured to achieve global predictive intelligence by collecting and analyzing large volumes of storage system telemetry data points to enable effortless management, analysis, and support. Indeed, such storage systems may be able to predict both capacity and performance, as well as generate intelligent advice for workload deployment, interaction, and optimization. Such AI platforms may be configured to scan all incoming storage system telemetry data against a library of problem fingerprints to capture hundreds of performance-related variables that are used to predict and resolve incidents and forecast performance loads in real time, before the incidents impact customer environments.

[0178] For further explanation, FIG. 4 illustrates a block diagram showing multiple storage systems (402, 404, 406) supporting pods in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (402, 404, 406) shown in FIG. 4 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage systems (402, 404, 406) shown in FIG. 4 may include the same components as the storage systems described above, fewer components, or additional components.

4, each of the storage systems (402, 404, 406) is shown as having at least one computer processor (408, 410, 412), computer memory (414, 416, 418), and computer storage (420, 422, 424). In some embodiments, the computer memory (414, 416, 418) and the computer storage (420, 422, 424) may be part of the same hardware device, while in other embodiments, the computer memory (414, 416, 418) and the computer storage (420, 422, 424) may be part of different hardware devices. The difference between computer memory (414, 416, 418) and computer storage (420, 422, 424) in this particular example is that computer memory (414, 416, 418) may be physically proximate to computer processors (408, 410, 412) and may store computer program instructions executed by computer processors (408, 410, 412), while computer storage (420, 422, 424) may be implemented as non-volatile storage for storing user data, metadata describing the user data, etc. For example, with reference to the above example of Figure 1A, the computer processors (408, 410, 412) and computer memory (414, 416, 418) for particular storage systems (402, 404, 406) may reside within one or more of the controllers (110A-110D). Meanwhile, the attached storage devices (171A to 171F) may function as computer storage (420, 422, 424) within a particular storage system (402, 404, 406).

In the example shown in FIG. 4, the illustrated storage systems (402, 404, 406) may be attached to one or more pods (430, 432) in accordance with some embodiments of the present disclosure. Each of the pods (430, 432) shown in FIG. 4 may include a dataset (426, 428). For example, a first pod (430) with three attached storage systems (402, 404, 406) includes a first dataset (426), and a second pod (432) with two attached storage systems (404, 406) includes a second dataset (428). In such an example, when a particular storage system attaches to a pod, the pod's dataset is copied to the particular storage system and then kept up to date as the dataset is modified. A storage system can be removed from a pod, resulting in the dataset no longer being kept up to date on the removed storage system. In the example shown in Figure 4, any storage system that is active for a pod (that is, it is the most recent, operational, non-faulted member of a non-faulted pod) can receive and process requests to modify or read the pod's dataset.

[0181] In the example shown in FIG. 4, each pod (430, 432) may include a collection of managed objects and management operations, as well as a collection of access operations to modify or read the dataset (426, 428) associated with the particular pod (430, 432). In such an example, a management operation may equally modify or query a managed object through either of the storage systems. Similarly, an access operation to read or modify a dataset may equally operate through either of the storage systems. In such an example, each storage system stores a separate copy of the dataset as an appropriate subset of the dataset stored and advertised for use by the storage system, but a managed object or operation to modify a dataset performed and completed through any one storage system is reflected in a subsequent managed object to query the pod or a subsequent access operation to read the dataset.

[0182] The reader will understand that pods may implement more functionality than simply clustered, synchronously replicated data sets. For example, pods can be used to implement tenants whereby data sets are securely isolated from one another in some way. Pods can also be used to implement virtual arrays or virtual storage systems where each pod presents as a unique storage entity on a network (e.g., a storage area network or Internet Protocol network) with a separate address. In the case of a multi-storage system pod implementing a virtual storage system, all physical storage systems associated with the pod may present themselves in some way as the same storage system (e.g., as if multiple physical storage systems were no different than multiple network ports into a single storage system).

[0183] The reader understands that a pod may be a unit of management representing a collection of volumes, file systems, object/analytic stores, snapshots, and other management entities, and that management changes in any one storage system (e.g., renaming, property changes, exporting, or managing permissions for any portion of the pod's dataset) are automatically reflected in all active storage systems associated with the pod. Furthermore, a pod may also be a unit of data collection and data analysis, where performance and capacity metrics may be presented aggregated across all active storage systems for the pod, or presented to challenge data collection and analysis per pod separately, or perhaps presenting the contribution of each attached storage system to incoming content and performance per pod.

[0184] One model of pod membership may be defined as a list of storage systems and a subset of that list of storage systems that are considered in sync for a pod. A storage system may be considered in sync for a pod if the pod is at least in recovery with identical idle content for the last read copy of the dataset associated with the pod. Idle content is the content after any ongoing modifications have completed without processing new modifications. This is sometimes referred to as "crash-recoverable" consistency. Pod recovery implements a process to reconcile differences in applying concurrent updates to synchronized storage systems within a pod. Recovery may resolve any inconsistencies between storage systems upon completion of concurrent modifications that were requested for various members of the pod but not signaled to either requestor as successfully completed. Storage systems that are listed as pod members but not listed as in sync for a pod can be described as "detached" from the pod. Storage systems listed as pod members are in sync for a pod and are currently available to actively serve data for the pod when they are "online" for the pod.

[0185] Each storage system member of a pod may have its own copy of the membership, including which storage systems it last knew were in sync and which storage systems it last knew contained the complete set of pod members. To be online for a pod, a storage system must consider itself in sync for the pod and be in communication with all other storage systems it considers in sync for the pod. If a storage system cannot be sure that it is in sync and in communication with all other storage systems that are in sync, then the storage system must stop processing new incoming requests for the pod (or complete them with an error or exception) until it is sure that it is in sync and in communication with all other storage systems that are in sync. A first storage system may determine that a second paired storage system should be detached, which allows the first storage system to continue because it is currently in sync with all systems in its list. Instead, however, the second storage system must be prevented from determining that the first storage system should be detached while the second storage system continues to operate. This would create a "split-brain" condition that could lead to incoherent datasets, dataset corruption, or application corruption, among other risks.

[0186] Situations requiring a determination of how to proceed when not communicating with a paired storage system may occur while a storage system is running normally and then realizes that communication has been lost, while it is currently recovering from some previous failure, while it is rebooting or resuming from a temporary power loss or restored communication outage, while it is switching operation from one set of storage system controllers to another for whatever reason, or during or after any combination of these or other types of events. In practice, whenever a storage system associated with a pod cannot communicate with all known non-detached members, the storage system either waits a short time until communication can be established, goes offline, and continues waiting, or the storage system may determine, by some means, that it is safe to detach the non-communicating storage system and then proceed, without risk of the non-communicating storage system inferring otherwise and causing a split-brain situation. If the safe detach occurs quickly enough, the storage system can remain online for the pod with little more than a short delay and without causing application outages for applications that may issue requests to the remaining online storage system.

[0187] An example of this situation is when a storage system may know it is out of date. This may occur, for example, when a first storage system is first added to a pod that is already associated with one or more storage systems, or when the first storage system reconnects to another storage system and finds that the other storage system has already marked the first storage system as detached. In this case, the first storage system simply waits until it connects to some other set of storage systems that it is in sync with for the pod.

[0188] This model requires some consideration of how storage systems are added to or removed from a pod or from the syncing pod member list. Because each storage system has its own copy of the list, and because two independent storage systems cannot update their local copies at exactly the same time, and because the local copies are all that is available on reboot or in various failure situations, care must be taken to ensure that transient inconsistencies do not cause problems. For example, if one storage system is syncing for a pod and a second storage system is added, and then the second storage system is updated to show both storage systems as syncing initially, then if there is a failure and a restart of both storage systems, the second may start up and wait to connect to the first storage system. Meanwhile, the first may not be aware that it should or will wait for the second storage system. If the second storage system then responds to its inability to connect with the first storage system by going through a process to detach it, it may then succeed in completing a process that the first storage system is unaware of, potentially resulting in a split-brain. Therefore, it may be necessary to ensure that storage systems do not inappropriately disagree as to whether they may choose to go through the detach process when they are not communicating.

[0189] One way to ensure that storage systems do not inappropriately disagree regarding whether they may choose to undergo the detach process if they are not communicating is to ensure that when adding a new storage system to the in-sync member list for a pod, the new storage system first remembers that it is a detached member (and possibly that it has been added as an in-sync member). Existing in-sync storage systems can then locally remember that the new storage system is an in-sync pod member before the new storage system remembers that same fact locally. If there is a reboot or network outage set up before the new storage system remembers its in-sync status, then the original storage system may detach the new storage system due to lack of communication, while the new storage system waits. To remove a communicating storage system from a pod, the reverse version of this change may be required. First, the removing storage system remembers that it is no longer in-sync, then the remaining storage systems remember that the removing storage system is no longer in-sync, and then all storage systems remove the removing storage system from their pod membership lists. Depending on the implementation, the intermediate persistent detach state may not be necessary. Whether or not local copies of membership lists require attention may depend on the model storage system to monitor each other or to validate its membership. If a consensus model is used for both, or if an external system (or external distributed or clustered system) is used to store and validate pod membership, then inconsistencies in locally stored membership lists may not be an issue.

[0190] When communication fails or one or more storage systems in a pod fail, or a storage system starts up (or fails over to a secondary controller) and is unable to communicate with its paired storage systems for a pod, and one or more storage systems decide to detach one or more paired storage systems, some algorithm or mechanism must be utilized to determine that it is safe to do so and to follow through on the detachment. One means of resolving detachment is to use a majority (or quorum) model for membership. In the case of three storage systems, as long as two are communicating, they can agree to detach the non-communicating third storage system, but that third storage system cannot independently choose to detach either of the other two. Confusion can arise when storage system communication is inconsistent. For example, storage system A may be communicating with storage system B but not with C. Meanwhile, storage system B may be communicating with both A and C. So A and B will detach C, or B and C will detach A, but realizing this may require more communication between pod members.

[0191] Care must be taken with the quorum membership model when adding and removing storage systems. For example, if a fourth storage system is added, then the "majority" of storage systems is now three. Transitioning from three storage systems (where two is required for a majority) to a pod containing the fourth storage system (where three is required for a majority) may require something similar to the model described above to carefully add the storage system to the synchronization list. For example, the fourth storage system starts in an attaching state but is not yet attached, which would never trigger a quorum vote. Once in that state, each of the original three pod members is updated to recognize the fourth member and the new requirement for a majority of three storage systems to detach the fourth member. Removing a storage system from a pod may similarly move that storage system to a locally stored "detached" state before updating other pod members. A variant of this is to use a distributed consensus mechanism, such as PAXOS or RAFT, to implement any membership changes or handle detachment requests.

[0192] Another means of managing membership transitions is to use an external system outside of the storage system itself to handle pod membership. To come online for a pod, the storage system must first contact the external pod membership system to verify that it is synchronized for the pod. Any storage system that is online for a pod should then remain in communication with the pod membership system and wait or go offline if the storage system loses communication. The external pod membership manager could be implemented as a highly available cluster using a variety of cluster tools, such as Oracle RAC, Linux HA, VERITAS Cluster Server, IBM's HACMP, or others. The external pod membership manager could also use a distributed configuration tool such as Etcd or Zookeeper, or a trusted distributed database such as Amazon's DynamoDB.

4, in accordance with some embodiments of the present disclosure, the illustrated storage systems (402, 404, 406) may receive a request to read a portion of a dataset (426, 428) and process the request to read the portion of the dataset locally. The reader will understand that because the dataset (426, 428) should be consistent across all storage systems (402, 404, 406) of the pod, a request (e.g., a write operation) to modify the dataset (426, 428) requires coordination among the storage systems (402, 404, 406) of the pod, but fulfilling a request to read a portion of the dataset (426, 428) does not require similar coordination among the storage systems (402, 404, 406). Thus, a particular storage system receiving a read request may service the read request locally by reading a portion of the dataset (426, 428) stored in the storage system's storage device, without synchronous communication with other storage systems in the pod. Read requests received by one storage system for a replicated dataset in a replicated cluster are expected to avoid any communication in the majority of cases, at least when received by a storage system running in the cluster that is also nominally running. Such reads should typically be successfully processed by simply reading from the local copy of the clustered data, with no additional interaction with other storage systems in the cluster required.

[0194] The reader understands that storage systems may take measures to ensure read consistency, so that read requests return the same result regardless of which storage system processes the read request. For example, the resulting clustered dataset contents for any set of updates received by any set of storage systems in a cluster should be consistent across the cluster, at least whenever the updates are idle (all previous modification operations are marked as complete, and no new update requests are being received or processed). More specifically, instances of a clustered dataset across a set of storage systems may differ only as a result of updates that have not yet been completed. That is, for example, any two write requests that overlap on that volume block range, or any combination of a snapshot, compare-and-write, or virtual block range copy that overlaps with a write request, must produce consistent results across all copies of the dataset. Two operations should not produce results as if they occurred in one order on one storage system and in a different order on another storage system of a replicated cluster.

[0195] Additionally, read requests may be made time-ordered consistent. For example, if one read request is received and completed at a replica cluster, and that read is then followed by another read request for an overlapping address range, and one or both reads overlap in time and volume address range with a modify request received by the replica cluster in any way (whether the read or the modify is received by the same storage system or a different storage system in the replica cluster), then if the first read reflects the results of the update, then the second read should also reflect the results of that update, rather than possibly returning data that preceded the update. If the first read does not reflect the update, then the second read will either reflect the update or not. This ensures that "time" for the data segment cannot progress backward between the two read requests.

[0196] In the example shown in FIG. 4, the illustrated storage systems (402, 404, 406) may detect an interruption in data communication with one or more of the other storage systems and determine whether the particular storage system should remain in the pod. The interruption in data communication with one or more of the other storage systems may occur for a variety of reasons. For example, the interruption in data communication with one or more of the other storage systems may occur because one of the storage systems has failed, because the network interconnect has failed, or for some other reason. An important aspect of a synchronized replicated cluster is ensuring that any failure handling does not result in irrecoverable discrepancies or any discrepancies in responses. For example, if the network fails between two storage systems, at most one of the storage systems can continue to process new incoming I/O requests for the pod. And, if one storage system continues processing, the other storage system will not be able to process any new requests, including read requests, to completion.

4, the illustrated storage systems (402, 404, 406) may determine whether a particular storage system should remain in a pod in response to detecting an interruption in data communication with one or more of the other storage systems. As described above, to be “online” as part of a pod, a storage system must consider itself in sync for the pod and be in communication with all other storage systems that it considers in sync for the pod. If a storage system cannot be confident that it is in sync and in communication with all other storage systems that are in sync, then the storage system stops processing new incoming requests to access the datasets (426, 428). Thus, a storage system may determine whether a particular storage system should remain online as part of a pod through, for example, a combination of both steps where the storage system must verify that it can communicate with all other storage systems it considers in-sync for the pod and that all other storage systems it considers in-sync for the pod also consider the storage system to be attached to the pod, by determining (via one or more test messages) whether the storage system can communicate with all other storage systems it considers in-sync for the pod, and that all other storage systems it considers in-sync for the pod also consider the storage system to be attached to the pod, or through some other mechanism.

4, the illustrated storage systems (402, 404, 406) may keep the datasets on the particular storage system accessible for management and dataset operations in response to determining that the particular storage system should remain within the pod. The storage system may keep the datasets (426, 428) on the particular storage system accessible for management and dataset operations, for example, by accepting and processing requests to access versions of the datasets (426, 428) stored on the storage system, by accepting and processing management operations associated with the datasets (426, 428) issued by a host or authorized administrator, by accepting and processing management operations associated with the datasets (426, 428) issued by one of the other storage systems, or in some other manner.

[0199] However, in the example shown in FIG. 4, the illustrated storage systems (402, 404, 406) may make datasets on the particular storage systems inaccessible for management and dataset operations in response to determining that the particular storage system should not remain in the pod. The storage systems may make datasets (426, 428) on the particular storage systems inaccessible for management and dataset operations, for example, by rejecting requests to access versions of the datasets (426, 428) stored on the storage systems, by rejecting management operations associated with the datasets (426, 428) issued by a host or other authorized administrator, by rejecting management operations associated with the datasets (426, 428) issued by one of the other storage systems in the pod, or in some other manner.

[0200] In the example shown in FIG. 4, the illustrated storage systems (402, 404, 406) may detect that an interruption in data communication with one or more of the other storage systems has been repaired and may make the datasets on the particular storage system accessible for management and dataset operations. The storage system may detect that the interruption in data communication with one or more of the other storage systems has been repaired, for example, by receiving a message from one or more of the other storage systems. In response to detecting that the interruption in data communication with one or more of the other storage systems has been repaired, the storage system may make the datasets (426, 428) on the particular storage system accessible for management and dataset operations once the previously detached storage system has resynchronized with the storage systems that remain attached to the pod.

4, the illustrated storage systems (402, 404, 406) may be taken offline from the pod such that the particular storage system no longer enables management and dataset operations. The illustrated storage systems (402, 404, 406) may be taken offline from the pod such that the particular storage system no longer enables management and dataset operations for a variety of reasons. For example, the illustrated storage systems (402, 404, 406) may be taken offline from the pod because of some failure at the storage system itself, because an update or some other maintenance is occurring at the storage system, because of a communication failure, or for many other reasons. In such an example, as described in more detail in the resynchronization section included below, the particular storage system may be taken offline and then brought back online with the pod such that the particular storage system enables management and dataset operations. The illustrated storage systems (402, 404, 406) may subsequently update the dataset on the particular storage system to include all updates to the dataset, as the particular storage system goes offline and comes back online with the pod such that the particular storage system enables management and dataset operations.

[0202] In the example shown in FIG. 4, the illustrated storage systems (402, 404, 406) may identify a target storage system to asynchronously receive a dataset, where the target storage system is not one of multiple storage systems to which the dataset is synchronously replicated. Such a target storage system may represent a backup storage system, for example, as any storage system that utilizes a synchronously replicated dataset. In practice, synchronous replication may be used to distribute copies of a dataset closer to some rack of servers for better local read performance. One such case is a smaller top-of-rack storage system that is symmetrically replicated to a larger storage system centrally located in a data center or campus, which is more carefully managed for reliability or connected to an external network for asynchronous replication or backup services.

[0203] In the example shown in FIG. 4, the illustrated storage systems (402, 404, 406) may identify a portion of a dataset that has not been asynchronously replicated to the target storage system by any of the other storage systems and asynchronously replicate to the target storage system the portion of the dataset that has not been asynchronously replicated to the target storage system by any of the other storage systems, with the two or more storage systems collectively replicating the entire dataset to the target storage system. In this manner, the work associated with asynchronously replicating a particular dataset may be divided among the members of the pod, such that each storage system of the pod is only responsible for asynchronously replicating a subset of the dataset to the target storage system.

[0204] In the example shown in FIG. 4, the illustrated storage systems (402, 404, 406) may be detached from the pod, such that the particular storage system detached from the pod is no longer included in the set of storage systems across which a dataset is synchronously replicated. For example, if storage system (404) in FIG. 4 were to detach from pod (430) shown in FIG. 4, pod (430) would only include storage systems (402, 406) as storage systems across which dataset (426) included in pod (430) would be synchronously replicated. In such an example, detaching a storage system from a pod might also include deleting the dataset from the particular storage system detached from the pod. Continuing with the example in which storage system (404) in FIG. 4 has detached from pod (430) shown in FIG. 4, dataset (426) included in pod (430) might be deleted or otherwise removed from storage system (404).

[0205] The reader will also understand that there are several unique management features enabled by the pod model that can be supported. The pod model itself also presents several issues that may be addressed by an implementation. For example, when a storage system is offline for a pod but is otherwise running because an interconnect failed and another storage system for the pod won arbitration, there may be a desire or need to access the offline pod's dataset on the offline storage system. One solution may be to simply enable the pod in some detached mode so that the dataset can be accessed. However, that solution is risky, and it may make it much more difficult to reconcile the pod's metadata and data when the storage systems regain communication. Furthermore, there will also be separate paths for the host to access the offline storage system as well as the storage system that is still online. In that case, the host may issue I/O to both storage systems even though they are no longer kept in sync. This is because the host sees target ports reporting a volume with the same identifier, and the host I/O driver infers that it sees additional paths to the same volume. Even if the host makes a naive guess, this can result in significant data corruption, as the reads and writes issued to both storage systems no longer match. A variant of this case is in a clustered application, such as a shared storage clustered database, where the clustered application running on one host might be reading from or writing to one storage system, and the same clustered application running on another host might be reading from or writing to a "detached" storage system, but the two instances of the clustered application communicate with each other assuming that the data sets they each see are perfectly consistent with respect to completed writes. Because they are not consistent, that assumption is violated, and the application's data set (e.g., a database) can quickly become corrupted.

[0206] One way to solve both of these problems is to allow offline pods, or perhaps snapshots of offline pods, to be copied to a new pod with a new volume that has a sufficiently new identity that host I/O drivers and clustered applications will not confuse the copied volume with a still-online volume on another storage system. Because each pod maintains a complete copy of the dataset that is crash consistent but perhaps slightly different from the copy of the pod dataset on another storage system, and because each pod has an independent copy of all data and metadata needed to operate on the pod contents, making a virtual copy of some or all of the pod's volumes or snapshots to a new volume in the new pod is a straightforward problem. For example, in a logical extent graph implementation, all that is required is to define a new volume in the new pod that references the pod's volumes or snapshots and the logical extent graph from the copied pod associated with the logical extent graph being marked as copied on write. Similar to how a volume snapshot that is copied to a new volume might be implemented, the new volume should be treated as a new volume. The volume may have the same administrative name, albeit in a new pod namespace. However, the volume should have a different fundamental identifier and a different logical unit identifier than the original volume.

[0207] In some cases, it may be possible to use virtual network isolation techniques (e.g., by creating a virtual LAN in the case of an IP network, or a virtual SAN in the case of a Fibre Channel network) so that isolation of volumes presented to some interfaces can be ensured to be inaccessible from host network interfaces or host SCSI initiation ports that might also see the original volume. In such cases, it may be safe to provide the copy of the volume with the same SCSI or other storage identifiers as the original volume. This could be used, for example, if an application expects to see a particular set of storage identifiers in order to function without undue burden on reconfiguration.

[0208] Some of the techniques described herein may also be used outside of active disaster situations to test readiness to handle disasters. Readiness testing (sometimes called "fire drills") is typically required for disaster recovery configurations, and frequent and repeated testing is deemed necessary to ensure that most or all aspects of the disaster recovery plan are correct and account for any recent changes to the application, data set, or equipment changes. Readiness testing should be non-disruptive to current production operations, including replicas. In many cases, the actual operations cannot actually be invoked in the active configuration, but a good approach is to use storage operations to make copies of the production data set, and then perhaps combine this with the use of virtual networking to create an isolated environment containing all the data deemed necessary for critical applications that must start successfully in the event of a disaster. Having such a copy of a synchronously replicated (or even asynchronously replicated) data set available at the site (or collection of sites) expected to perform disaster recovery readiness testing procedures and then start critical applications on that data set to ensure it starts and functions is an excellent tool, as it helps ensure that critical portions of the application data set were not left out of the disaster recovery plan. Where necessary and practical, this could be combined with a virtual isolation network, perhaps coupled with an isolated collection of physical or virtual machines, to approximate as closely as possible a real-world disaster recovery takeover situation. Virtually copying a pod (or collection of pods) to another pod as a point-in-time image of the pod dataset not only allows isolation to a single site (or several sites) separate from the original pod, but also immediately creates an isolated dataset containing all the copied elements and that can be operated essentially the same as the original pod. Furthermore, these are fast operations, and they can be easily torn down and tested repeatedly as often as desired.

[0209] Several enhancements could be made to move even closer toward a complete disaster recovery test. For example, in conjunction with isolated networks, SCSI logical unit identities or other types of identities could be copied to the target pod so that test servers, virtual machines, and applications see the same identifiers. Furthermore, the server management environment would be configured to respond to requests from a specific virtual set of virtual networks and respond to requests or operations on the original pod name so that scripts do not require the use of test variants along with alternate "test" versions of object names. Additional enhancements could be used when the hosting server infrastructure that takes over in the event of a disaster is available for use during testing. This includes cases where a disaster recovery datacenter is generally fully equipped with alternate server infrastructure that is not used until dictated to do so by a disaster. It also includes cases where the infrastructure may be used for non-critical operations (e.g., performing analytics on production data, or simply supporting application development or other functions that may be critical but can be paused if needed for more critical functions). Specifically, host definitions and configurations, and the server infrastructure that uses them, can be set up as they are for an actual disaster recovery takeover event, tested as part of the disaster recovery takeover test, with the tested volumes connected to these host definitions from the virtual pod copies used to provide snapshots of the datasets. From the perspective of the storage systems involved, these host definitions and configurations used for testing, as well as the volume-to-host connection configurations used during testing, can then be reused when an actual disaster takeover event is triggered, significantly minimizing the configuration differences between the test configuration and the actual configuration that will be used in the event of a disaster recovery takeover.

[0210] In some cases, it may make sense to move volumes from a first pod and to a new, second pod containing only those volumes. Pod membership and high availability and recovery features can then be adjusted separately, and management of the two resulting pod data sets can then be decoupled from each other. Also, operations that can be performed in one direction should also be possible in the other direction. At some point, it may make sense to take two pods and merge them into one, such that the volumes in each of the original two pods now track each other for storage system membership and high availability and recovery features and events. Both operations can be achieved safely and with reasonably minimal or no disruption to running applications by relying on the proposed features for changing the brokering or quorum characteristics for pods, as described in the sections above. For example, through mediation, a pod's mediator can be modified using a sequence consisting of steps in which each storage system in the pod is modified to depend on both a first mediator and a second mediator, and then each is modified to depend only on the second mediator. If a failure occurs midway through the sequence, some storage systems may depend on both the first mediator and the second mediator, but recovery and failure handling never results in some storage systems depending only on the first mediator and other storage systems depending only on the second mediator. Quorum can be handled similarly by temporarily relying on winning both the first and second quorum models to proceed to recovery. While this may result in a very brief period in which the availability of a pod facing a failure depends on additional resources, thus reducing potential availability, this period is so brief that the reduction in availability is often negligible. With intermediation, if the change in mediator parameters is merely a change in the key used for intermediation and the intermediary service used is the same, then the potential reduction in availability is also less, as it now relies on two calls to the same service versus one call to that service, and only on two separate services rather than separate calls.

[0211] The reader should note that changing the quorum model can be quite complicated. Additional steps may be required in which the storage system joins the second quorum model but does not rely on winning in that second quorum model, followed by steps that also rely on the second quorum model. This may be necessary to account for the fact that if only one system processes the change to rely on the quorum model, then it will never achieve quorum because there will never be a majority. With this model in place to change high availability parameters (intermediary relationships, quorum model, takeover priority), safety procedures can be created for splitting a pod into two or joining two pods into one. This may require adding one other feature: linking a second pod to a first pod for high availability, so that if the two pods contain compatible high availability parameters, the second pod linked to the first pod can rely on the first pod to determine and drive detach relationship handling and behavior, offline and sync states, and recovery and resync actions.

[0212] To split one pod into two, an operation for moving some volumes into a newly created pod may be performed, which can be described as follows: creating a second pod that moves a collection of volumes that were previously in the first pod; copying high availability parameters from the first pod to the second pod to ensure they are compatible for linking; and linking the second pod to the first pod for high availability. This operation may be encoded as a message and should be implemented by each storage system in the pod, such that the storage system ensures that the operation occurs entirely on that storage system, or not at all if processing is interrupted by a failure. Once all synchronized storage systems for the two pods have processed this operation, the storage system may then process a subsequent operation that modifies the second pod so that it is no longer linked to the first pod. As with other changes to high availability features for pods, this involves first having each in-sync storage system rely on both the previous model (where high availability is linked to the first pod) and the new model (where it is now highly available independently). In the case of brokering or quorum, this means that the storage system that processed the change first relies on the brokering or quorum achieved accordingly for the first pod, and then relies on the new, separate brokering (e.g., new broker key) or quorum being achieved for the second pod before proceeding following the failure that required the test for brokering or quorum. As with the above description of changing the quorum model, an intermediate step may be configuring the storage system to participate in a quorum for the second pod before the step in which the storage system joins and relies on the quorum for the second pod. Once all in-sync storage systems have processed the change to rely on the new parameters for brokering or quorum for both the first pod and the second pod, the split is complete.

[0213] Joining a second pod to a first pod essentially works in reverse. First, the second pod must be adjusted to be compatible with the first pod by having an identical list of storage systems and a compatible high availability model. This may involve any set of steps, such as steps for adding or removing storage systems or for modifying mediator and quorum models, as described elsewhere herein. Depending on the implementation, it may only be necessary to arrive at an identical list of storage systems. Joining proceeds by processing operations at each synchronizing storage system to link the second pod to the first pod for high availability. Each storage system that processes the operation then depends on the first pod for high availability and then on the second pod for high availability. Once all synchronized storage systems for the second pod have processed that operation, the storage system then processes subsequent operations to remove the link between the second pod and the first pod, migrate the volume from the second pod to the first pod, and delete the second pod, respectively. Host or application data set access can be preserved throughout these operations, so long as the implementation allows appropriate direction of host or application data set modification or read operations to volumes by identity, and so long as identity is preserved appropriately for the storage protocol or storage model (e.g., as long as logical unit identifiers for volumes and use of target ports to access volumes are preserved in the case of SCSI).

[0214] Migrating volumes between pods can present problems. If the pods have the same set of in-sync membership storage systems, then it may be simple to temporarily suspend operations on the migrating volumes, switch control over operations on those volumes to the control software and structures for the new pod, and then resume operations. If networks and ports migrate appropriately between pods, this allows for a seamless transition with continuous uptime for applications, apart from a very brief interruption in operations. Depending on the implementation, the suspend operation may not be necessary, or may be so internal to the system that the interruption in operations has no impact. Copying volumes between pods with different in-sync membership sets is increasingly problematic. This is less of an issue if the target pod of the copy has a subset of in-sync members from the source pod. Member storage systems can be removed safely enough without the need for more work. However, if the target pod adds an in-sync member storage system to a volume on the source pod, then the added storage system must be synchronized to include the volume's contents before they can be used. Until synchronized, this leaves the copied volumes distinctly different from volumes that are already synchronized, in that failure handling is different and requests from member storage systems that are not yet synchronized may not work or may have to be forwarded or may not be fast because reads must traverse the interconnect. Also, the internal implementation must handle some volumes being synchronized and ready for failure handling, while others are not.

[0215] There are other issues related to the reliability of operations in the face of failures. Coordinating the migration of volumes across multiple storage system pods is a distributed operation. If the pod is the unit of failure handling and recovery, and if intermediary or quorum or whatever means are used to avoid split-brain situations, then switching a volume from one pod to another pod, and then to the pod's storage system, with a particular set of state, configuration, and relationships for failure handling, recovery, intermediary, and quorum, must be careful to coordinate the changes involved in that operation for every volume. Operations cannot be atomically distributed across storage systems, but must be phased in some way. The intermediary and quorum models inherently provide tools for implementing distributed transaction atomicity within pods, but this may not extend to inter-pod operations without increasing implementation complexity.

[0216] Consider also the simple migration of a volume from a first pod to a second pod in the case of two pods sharing the same first and second storage systems. At some point, the storage systems coordinate to determine that the volume is now in the second pod and no longer in the first pod. Without an inherent mechanism for transaction atomicity across storage systems for the two pods, a native implementation might leave the volume in the first pod on the first storage system and the second pod on the second storage system in the event of a network failure that triggers a failure process to detach the storage systems from the two pods. If the pods independently determine which storage system succeeds in detaching the other, the result would be either the same storage system detaching the other for both pods, in which case the volume migration recovery results must be consistent, or different storage systems would detach the other for the two pods. If a first storage system detaches a second storage system for a first pod, and the second storage system detaches the first storage system for the second pod, then recovery will recover the volume to the first pod on the first storage system and to the second pod on the second storage system, and the volume will then be running and exported to hosts and storage applications on both storage systems. Alternatively, if a second storage system detaches the first storage system for the first pod, and the first storage system detaches the second storage system for the second pod, then recovery may result in the volume being discarded from the second pod by the first storage system and the volume being discarded from the first pod by the second storage system, causing the volume to disappear entirely. If the pods between which the volume is being migrated are on different sets of storage systems, then things may become even more complicated.

[0217] A solution to these problems may be to use an intermediate pod in conjunction with the techniques described above for splitting and joining pods. This intermediate pod must never be presented as a visible management object associated with the storage system. In this model, a volume to be moved from a first pod to a second pod is first split from the first pod into a new intermediate pod using the split operation described above. The storage system membership for the intermediate pod can then be adjusted to match the membership of the storage system by adding or removing storage systems from the pod as needed. The intermediate pod may then be joined to the second pod.

[0218] For additional explanation, FIG. 5 sets forth a flowchart illustrating steps that may be performed by storage systems (402, 404, 406) supporting pods in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (402, 404, 406) illustrated in FIG. 5 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, 4, or any combination thereof. In practice, the storage systems (402, 404, 406) illustrated in FIG. 5 may include the same components as the storage systems described above, fewer components, or additional components.

[0219] In the example method shown in FIG. 5, a storage system (402) may be attached to a pod (508). The model for pod membership may include a list of storage systems and a subset of that list of storage systems that are presumed to be in sync for the pod. A storage system is in sync for a pod if it is in recovery with at least the same idle content of the last created copy of the dataset associated with the pod. Idle content is the content after no new modifications are processed and any ongoing modifications have completed. This is sometimes referred to as "crash recoverable" consistency. A storage system that is listed as a pod member but not listed as in sync for the pod may be described as "detached" from the pod. A storage system that is listed as a pod member is in sync for the pod and is currently available to actively serve data for the pod when it is "online" for the pod.

5, the storage system (402) may attach (508) to the pod by synchronizing its locally stored version of the dataset (426) with the latest version of the dataset (426) stored in the pod's other storage systems (404, 406) that are online, e.g., as the terminology is described above. In such an example, in order for the storage system (402) to attach (508) to the pod, the pod definition stored locally in each of the storage systems (402, 404, 406) in the pod needs to be updated in order for the storage system (402) to attach (508) to the pod. In such an example, each storage system member of the pod may have its own copy of the membership, including which storage systems it last knew were in sync and which storage systems it last knew contained the entire set of pod members.

5, storage system (402) may receive a request to read a portion of dataset (426) (510), and storage system (402) may process the request to read the portion of dataset (426) locally (512). The reader will understand that while a request to modify dataset (426) (e.g., a write operation) requires coordination among storage systems (402, 404, 406) in the pod, servicing a request to read a portion of dataset (426) does not require similar coordination among storage systems (402, 404, 406) because dataset (426) must be consistent across all storage systems (402, 404, 406) in the pod. Thus, a particular storage system (402) receiving a read request may service the read request locally by reading a portion of the dataset (426) stored in its (402) storage device, without synchronous communication with other storage systems (404, 406) in the pod. Read requests received by one storage system for a replicated dataset in a replicated cluster are expected to avoid any communication in the vast majority of cases, at least when received by a storage system that is also running in a nominally running cluster. Such reads should typically be successfully processed by simply reading from the local copy of the clustered dataset, with no additional interaction with other storage systems in the cluster required.

[0222] The reader understands that storage systems may take steps to ensure read consistency, so that a read request returns the same result regardless of which storage system processes the read request. For example, the resulting clustered dataset contents for any set of updates received by any set of storage systems in a cluster should be consistent across the cluster, at least at any point in time when the updates are idle (all previous modification operations are marked as completed, and no new update requests have been received or processed). That is, instances of a clustered dataset across a set of storage systems may differ only as a result of updates that have not yet completed. That is, for example, any two write requests that overlap on that volume block range, or any combination of a snapshot, compare-and-write, or virtual block range copy that overlaps with a write request, must produce consistent results for all copies of the dataset. Two operations cannot produce results as if they occurred in one order on one storage system and in a different order on another storage system in a replication cluster.

[0223] Furthermore, read requests may be time-consistent. For example, if one read request is received and completed at a replicated cluster, and then that read is followed by another read request for an overlapping address range that is received by the replicated cluster, and one or both reads overlap in time and volume address range to any extent with the modify request received by the replicated cluster (regardless of whether the read or the modify is received by the same storage system in the replicated cluster or a different storage system), then if the first read reflects the results of the update, then the second read should also reflect the results of that update, presumably rather than returning data that preceded the update. If the first read does not reflect the update, then the second read may or may not reflect the update. This ensures that the "time" of the data segment never goes back in time between the two read requests.

5, the storage system (402) may detect (514) an interruption in data communication with one or more of the other storage systems (404, 406). The interruption in data communication with one or more of the other storage systems (404, 406) may occur for a variety of reasons. For example, the interruption in data communication with one or more of the other storage systems (404, 406) may occur because one of the storage systems (402, 404, 406) has failed, because the network interconnect has failed, or for some other reason. An important aspect of synchronous replication clustering is ensuring that any failure handling does not result in an irrecoverable inconsistency or correspondence inconsistency. For example, if the network fails between two storage systems, at most one of the storage systems may continue to process new incoming I/O requests for the pod. And, if one storage system continues processing, the other storage system will not be able to process new requests, including read requests, to completion.

5, a storage system (402) may determine whether a particular storage system (402) should remain online as part of a pod (516). As described above, to be 'online' as part of a pod, a storage system must consider itself in sync for the pod and be in communication with all other storage systems that it considers in sync for the pod. If a storage system cannot be sure that it is in sync and in communication with all other storage systems that are in sync, then the storage system may stop processing new incoming requests to access the dataset (426). Thus, the storage system (402) may determine whether a particular storage system (402) should remain online as part of a pod by determining (516) whether the storage system can communicate (e.g., via one or more test messages) with all other storage systems (404, 406) that it considers in-sync for the pod, by determining whether all other storage systems (404, 406) that it considers in-sync for the pod also consider the storage system (402) to be attached to the pod, by verifying that the particular storage system (402) can communicate with all other storage systems (404, 406) that it considers in-sync for the pod and that all other storage systems (404, 406) that it considers in-sync for the pod also consider the storage system (402) to be attached to the pod, or by some other mechanism.

5, in response to affirmatively determining (518) that a particular storage system (402) should remain online as part of a pod, the storage system (402) may keep (522) the dataset (426) on the particular storage system (402) accessible for management and dataset operations. The storage system (402) may make (522) the dataset (426) on the particular storage system (402) accessible for management and dataset operations by, for example, accepting and processing requests to access versions of the dataset (426) stored on the storage system (402), accepting and processing management operations associated with the dataset (426) issued by a host or authorized administrator, accepting and processing management operations associated with the dataset (426) issued by one of the other storage systems (404, 406) in the pod, or in some other manner.

5, the storage system (402) may also make the dataset (426) on the particular storage system (402) inaccessible for management and dataset operations (524) in response to determining (520) that the particular storage system should not remain online as part of the pod. The storage system (402) may make the dataset (426) on the particular storage system (402) inaccessible for management and dataset operations (524), for example, by rejecting requests to access the version of the dataset (426) stored on the storage system (402), by rejecting management operations associated with the dataset (426) issued by a host or other authorized administrator, by rejecting management operations associated with the dataset (426) issued by one of the other storage systems (404, 406) in the pod, or in some other manner.

5, the storage system (402) may detect (526) that a disruption in data communication with one or more of the other storage systems (404, 406) has been repaired. The storage system (402) may detect (526) that a disruption in data communication with one or more of the other storage systems (404, 406) has been repaired, for example, by receiving a message from one or more of the other storage systems (404, 406). In response to detecting (526) that a disruption in data communication with one or more of the other storage systems (404, 406) has been repaired, the storage system (402) may make (528) the dataset (426) on the particular storage system (402) accessible for management and dataset operations.

5 describes an embodiment in which various actions are shown as occurring in the same order, although no ordering is required. Additionally, other embodiments may exist in which the storage system (402) performs only a subset of the described actions. For example, the storage system (402) may perform the steps of: detecting (514) an interruption in data communication with one or more of the other storage systems (404, 406); determining (516) whether the particular storage system (402) should remain in the pod; keeping (522) a dataset (426) on the particular storage system (402) accessible for management and dataset operations; or making (524) a dataset (426) on the particular storage system (402) inaccessible for management and dataset operations without first receiving (510) a request to read a portion of the dataset (426) and processing (512) the request to read the portion of the dataset (426) locally. Additionally, the storage system (402) may detect (526) that an interruption in data communication with one or more of the other storage systems (404, 406) has been repaired and may make (528) the dataset (426) on the particular storage system (402) accessible for management and dataset operations without first receiving (510) a request to read a portion of the dataset (426) and processing (512) the request to read the portion of the dataset (426) locally. Indeed, none of the steps described herein are explicitly required in all embodiments as a prerequisite for performing any other step described herein.

[0230] For additional explanation, FIG. 6 sets forth a flowchart illustrating steps that may be performed by storage systems (402, 404, 406) supporting pods in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (402, 404, 406) illustrated in FIG. 6 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, 4, or any combination thereof. In practice, the storage systems (402, 404, 406) illustrated in FIG. 6 may include the same components as the storage systems described above, fewer components, or additional components.

6, two or more of the storage systems (402, 404) may each identify a target storage system (618) for asynchronously receiving the dataset (426) (608). The target storage system (618) for asynchronously receiving the dataset (426) may be implemented as cloud storage provided by a cloud service provider or in many other ways, such as a backup storage system located in a different data center than either of the storage systems (402, 404) that are members of a particular pod. The reader will understand that the target storage system (618) is not one of multiple storage systems (402, 404) across which the dataset (426) is synchronously replicated, and therefore the target storage system (618) does not initially contain an up-to-date local copy of the dataset (426).

6, two or more of the storage systems (402, 404) may each identify (610) a portion of the dataset (426) that has not been asynchronously replicated to a target storage system (618) by any of the other storage systems that are members of the pod that includes the dataset (426). In such an example, the storage systems (402, 404) may each asynchronously replicate (612) to the target storage system (618) a portion of the dataset (426) that has not been asynchronously replicated to the target storage system by any of the other storage systems. Consider an example in which a first storage system (402) is responsible for asynchronously replicating a first portion of the dataset (426) (e.g., a first half of its address space) to the target storage system (618). In such an example, the second storage system (404) would be responsible for asynchronously replicating a second portion of the data set (426) (e.g., a second half of the address space) to the target storage system (618) such that the two or more storage systems (402, 404) collectively replicate the entire data set (426) to the target storage system (618).

[0233] The reader will understand that, as described above, through the use of pods, the replication relationship between two storage systems can be switched from one in which data is asynchronously replicated to one in which data is synchronously replicated. For example, if storage system A is configured to asynchronously replicate a dataset to storage system B, creating a pod that includes the dataset, storage system A as a member, and storage system B as a member can switch the relationship from one in which data is asynchronously replicated to one in which data is synchronously replicated. Similarly, through the use of pods, the replication relationship between two storage systems can be switched from one in which data is synchronously replicated to one in which data is asynchronously replicated. For example, if a pod that includes a dataset, storage system A as a member, and storage system B as a member is created by simply not extending the pod (to remove storage system A as a member or to remove storage system B as a member), the relationship in which data is synchronously replicated between the storage systems can immediately be switched to one in which data is asynchronously replicated. In this manner, storage systems can be switched back and forth between asynchronous replication and synchronous replication as needed.

[0234] This switch may be facilitated by an implementation relying on similar techniques for both synchronous and asynchronous replication. For example, if resynchronization of synchronously replicated datasets relies on the same or compatible mechanisms used for asynchronous replication, then switching to asynchronous replication is conceptually identical to dropping the in-sync state and leaving the relationship in a state similar to "durable recovery" mode. Similarly, switching from asynchronous replication to synchronous replication may conceptually work by completing resynchronization and "catching up" as the switching system becomes an in-sync pod member, just as it does.

[0235] Alternatively, or in addition, if both synchronous and asynchronous replication rely on similar or identical common metadata, or a common model for representing and identifying logical extents or stored block identities, or a common model for representing content-addressable stored blocks, then these aspects of commonality may be exploited to dramatically reduce the content that may need to be transferred when switching to and from synchronous and asynchronous replication. Furthermore, if a data set is asynchronously replicated from storage system A to storage system B, and system B further asynchronously replicates that data set to storage system C, then the common metadata model, common logical extents or block identities, or common representation of content-addressable storage blocks can dramatically reduce the data transfers required to enable synchronous replication between storage system A and storage system C.

[0236] The reader will further understand that, by using pods as described above, replication techniques can be used to perform tasks other than data replication. Indeed, since a pod may include a collection of managed objects, tasks such as migrating a virtual machine can be performed using the pod and replication techniques described herein. For example, if virtual machine A is running on storage system A, by creating a pod that includes virtual machine A as a managed object, storage system A as a member, and storage system B as a member, virtual machine A and any associated images and definitions may be migrated to storage system B, at which time the pod may simply be destroyed, membership may be updated, or other action may be taken as necessary.

[0237] For further explanation, FIG. 7 sets forth a flowchart illustrating an example method for establishing a synchronous replication relationship between two or more storage systems (714, 724, 728) according to some embodiments of the present disclosure. While not shown in great detail, the storage systems (714, 724, 728) illustrated in FIG. 7 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage systems (714, 724, 728) illustrated in FIG. 7 may include the same components as the storage systems described above, fewer components, or additional components.

7 includes identifying (702), for a dataset (712), multiple storage systems (714, 724, 728) across which the dataset (712) is synchronously replicated. The dataset (712) shown in FIG. 7 may be embodied, for example, as the contents of a particular volume, as the contents of a particular shard of a volume, or as any other collection of one or more data elements. The dataset (712) may be synchronized across the multiple storage systems (714, 724, 728) such that each storage system (714, 724, 728) maintains a local copy of the dataset (712). In the example described herein, such a dataset (712) is synchronously replicated across storage systems (714, 724, 728) such that the dataset (712) can be accessed through any of the storage systems (714, 724, 728) that have performance characteristics such that any one storage system in the cluster operates substantially less optimally than any other storage system in the cluster, at least as long as the cluster and the particular storage system being accessed are nominally operating. In such a system, modifications to the dataset (712) should be made to copies of the dataset resident on each storage system (714, 724, 728) so that accessing the dataset (712) on any of the storage systems (714, 724, 728) produces consistent results. For example, a write request issued to the dataset must be serviced by all storage systems (714, 724, 728) or by none of the storage systems (714, 724, 728). Similarly, some groups of operations (e.g., two write operations directed to the same location in a dataset) must be performed in the same order on all storage systems (714, 724, 728) so that the copies of the dataset resident on each storage system (714, 724, 728) are ultimately identical. Modifications to dataset (712) need not occur at exactly the same time, although some actions (e.g., issuing an acknowledgment that a write request is directed to the dataset and allows read access to locations in the dataset targeted by the write request that have not yet completed on all storage systems) may be deferred until the copies of dataset (712) on each storage system (714, 724, 728) have been modified.

7, identifying (702), for a dataset (712), multiple storage systems (714, 724, 728) across which the dataset (712) is synchronously replicated may be performed, for example, by examining a pod definition or similar data structure that associates the dataset (712) with one or more storage systems (714, 724, 728) that nominally store the dataset (712). A 'pod,' as the term is used here and throughout the remainder of this application, may be implemented as a management entity representing a dataset, a collection of management objects and management operations, a collection of access operations to modify or read the dataset, and multiple storage systems. Such management operations may modify or query managed objects equally through any of the storage systems, with access operations to read or modify the dataset operating equally through any of the storage systems. Each storage system may store a separate copy of the dataset as an appropriate subset of the dataset stored and advertised for use by the storage system, and operations that modify a management object or dataset performed and completed through any one storage system are reflected in subsequent management objects to query the pod or subsequent access operations to read the dataset. Additional details regarding "pods" may be found in previously filed provisional patent application Ser. No. 62/518,071, which is incorporated herein by reference. In such an example, a pod definition may include at least an identification of a dataset (712) and a collection of storage systems (714, 724, 728) across which the dataset (712) is synchronously replicated. Such a pod may encapsulate some of many (possibly optional) characteristics, including symmetric access, flexible addition/removal of replicas, highly available data consistency, centralized user management across storage systems in relation to the dataset, management host access, application clustering, etc. A storage system may be added to a pod, resulting in the pod's dataset (712) being copied to that storage system and then kept up to date as the dataset (712) is modified. A storage system may also be removed from a pod, resulting in the dataset (712) being no longer kept up to date with the removed storage system. In such examples, the pod definition or similar data structure may be updated as storage systems are added to and removed from particular pods.

7 also includes configuring (704) one or more data communication links (716, 718, 720) between each of the multiple storage systems (714, 724, 728) used to synchronously replicate the data set (712). In the example method shown in FIG. 6, the storage systems (714, 724, 728) within a pod must communicate with each other both for high-bandwidth data transfers and for cluster communication, status communication, and management communication. These distinct types of communication may be over the same data communication links (716, 718, 720), or in alternative embodiments, these distinct types of communication may be over separate data communication links (716, 718, 720). In a cluster of dual-controller storage systems, both controllers of each storage system should have the nominal ability to communicate with both controllers for any paired storage system (i.e., any other storage system within the pod).

[0241] In a primary/secondary controller design, all cluster communications for active replication may occur between primary controllers until a failure occurs. In such systems, some communication may occur between the primary and secondary controllers, or between secondary controllers of separate storage systems, to verify that the data communication links between such entities are operational. In other cases, virtual network addresses may be used to limit the configuration required for network links between data centers or to simplify the design of clustered aspects of storage systems. In an active/active controller design, cluster communications may continue from all active controllers of one storage system to some or all active controllers of any paired storage system, or the cluster communications may be filtered through a common switch, or the cluster communications may use virtual network addresses to simplify configuration, or the cluster communications may use some combination. In a scale-out design, two or more common network switches may be used, with all scale-out storage controllers in the storage system connecting to the network switch for handling data traffic. The switch may or may not use techniques to limit the number of exposed network addresses, so that the paired storage system does not need to be configured with the network addresses of all storage controllers.

7, configuring (704) one or more data communication links (716, 718, 720) between each of the multiple storage systems (714, 724, 728) used to synchronously replicate the data set (712) may be performed, for example, by configuring the storage systems (716, 718, 720) to communicate via defined ports over a data communication network, by configuring the storage systems (716, 718, 720) to communicate via a point-to-point data communication link between two of the storage systems (716, 724, 728), or in a variety of other ways. If secure communication is required, some form of key exchange may be required, or communication may occur or be initiated through some service, such as SSH (Secure Shell), SSL, or some other service or protocol built around public key or Diffie-Hellman key sharing, or a suitable alternative. Secure communications could also be mediated through some vendor-provided cloud service that is tied in some way to the customer identity. Alternatively, a service configured to run on the customer premises, e.g., running on a virtual machine or virtual container, could be used to broker the key exchange necessary for secure communications between the replicating storage systems (716, 718, 720). The reader will understand that a pod containing more than two storage systems may require communication links between most or all of the individual storage systems. In the example shown in FIG. 6, three data communication links (716, 718, 720) are shown, although additional data communication links may be present in other embodiments.

[0243] The reader will understand that communication between the storage systems (714, 724, 728) across which the dataset (712) is synchronously replicated serves several purposes. For example, one purpose is to deliver data from one storage system (714, 724, 728) to another storage system (714, 724, 728) as part of an I/O operation. For example, processing a write typically requires delivering the write content and some description of the write to any paired storage system for a pod. Another purpose served by data communication between the storage systems (714, 724, 728) may be to communicate configuration changes and analysis data for processing the creation, expansion, deletion, or renaming of volumes, files, object buckets, etc. Another purpose served by data communication between the storage systems (714, 724, 728) may be to facilitate communications involved in detecting and handling storage system and interconnect failures. This type of communication can be time-critical and may need to be prioritized to ensure it doesn't get stuck behind long network queuing delays when large bursts of write traffic are suddenly dumped on the data center interconnect.

[0244] The reader will further understand that different types of communications may use the same or different connections, in various combinations, and may use the same or different networks. Furthermore, some communications may be encrypted and protected, while other communications may not be encrypted. In some cases, a data communications link could be used to transfer I/O requests from one storage system to another (either directly as the request itself or as a logical description of the operation that the I/O request represents). This could be used, for example, when one storage system has up-to-date, in-sync content for a pod and another storage system does not currently have up-to-date, in-sync content for the pod. In such a case, as long as the data communications link is running, a request can be transferred from a storage system that is not up-to-date and in-sync to a storage system that is up-to-date and in-sync.

7 also includes exchanging (706) timing information (710, 722, 726) between the plurality of storage systems (714, 724, 728) for at least one of the plurality of storage systems (714, 724, 728). In the example method shown in FIG. 6, the timing information (710, 722, 726) for a particular storage system (714, 724, 728) may be embodied, for example, as a value of a clock within the storage system (714, 724, 728). In an alternative embodiment, the timing information (710, 722, 726) for a particular storage system (714, 724, 728) may be embodied as a value that serves as a proxy for the clock value. The value that serves as a proxy for the clock value may be included in a token exchanged between the storage systems. For example, a value that acts as a proxy for a clock value may be implemented, such as a sequence number that a particular storage system (714, 724, 728) or storage controller may internally record as having been sent at a particular time. In such an example, if a token (e.g., a sequence number) is also received, the associated clock value may be found and utilized as a criterion for determining whether a valid lease is still in effect. In the example method shown in FIG. 6 , exchanging (706) timing information (710, 722, 726) among the plurality of storage systems (714, 724, 728) for at least one of the plurality of storage systems (714, 724, 728) may be performed, for example, by each storage system (714, 724, 728) sending timing information to each other storage system (714, 724, 728) in the pod periodically, on-demand, within a predetermined amount of time after lease establishment, within a predetermined amount of time before the lease is set to expire, as part of an attempt to initiate or re-establish a synchronous replication relationship, or in some other manner.

7 also includes establishing (708) a synchronous replication lease according to the timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728), the synchronous replication lease identifying a period of time during which the synchronous replication relationship is valid. In the example method shown in FIG. 7, the synchronous replication relationship is formed as a collection of storage systems (714, 724, 728) that replicate some data (712) between these largely independent stores, with each storage system (714, 724, 728) having its own copy and its own separate internal management of relevant data structures for defining storage objects, mapping objects to physical storage, for deduplication, defining mapping of content to snapshots, etc. Synchronous replication relationships may be specific to a particular data set, such that a particular storage system (714, 724, 728) may be associated with more than one synchronous replication relationship, each differentiated by the data set being described and may further comprise a different set of additional member storage systems.

7, a synchronous replication lease may be established (708) depending on timing information (710, 722, 726) for at least one of a plurality of storage systems (714, 724, 728) in a variety of different ways. In one embodiment, the storage system may establish a synchronous replication lease (708) by utilizing timing information (710, 722, 726) for each of the plurality of storage systems (714, 724, 728) to adjust the clock. In such an example, once the clock is adjusted for each of the storage systems (714, 724, 728), the storage system may establish a synchronous replication lease (708) that extends for a predetermined period beyond the adjusted clock value. For example, if the clock for each storage system (714, 724, 728) is adjusted to a value of X, the storage systems (714, 724, 728) may each be configured to establish a synchronous replication lease that is valid for up to X+2 seconds.

[0248] In an alternative embodiment, the need to coordinate clocks between storage systems (714, 724, 728) may be avoided while still achieving timing guarantees. In such an embodiment, the storage controller in each storage system (714, 724, 728) may have a local monotonically increasing clock. A synchronous replication lease may be established between storage controllers (e.g., a primary controller of one storage system communicating with a primary controller of a paired storage system) by each controller sending its clock value to the other storage controller along with the last clock value it received from the other storage controller (708). When a particular controller also receives its clock value from another controller, the particular controller adds any agreed-upon lease interval to the received clock value and uses it to establish its local synchronous replication lease (708). In this manner, the synchronous replication lease may be calculated according to the value of the local clock received from the other storage system.

[0249] Consider an example in which a storage controller of a first storage system (714) is communicating with a storage controller of a second storage system (724). In such an example, assume that the value of the monotonically increasing clock for the storage controller of the first storage system (714) is 1000 milliseconds. Further assume that the storage controller of the first storage system (714) sends a message to the storage controller of the second storage system (724) indicating that the value of its clock was 1000 milliseconds when the message was generated. In such an example, assume that 500 milliseconds after the storage controller of the first storage system (714) sends a message to the storage controller of the second storage system (724) indicating that its clock value was 1000 milliseconds when the message was generated, the storage controller of the first storage system (714) receives a message from the storage controller of the second storage system (724) indicating that 1) the value of the monotonically increasing clock of the storage controller of the second storage system (724) was 5000 milliseconds when the message was generated, and 2) the last value of the monotonically increasing clock of the storage controller of the first storage system (714) received by the second storage system (724) was 1000 milliseconds. In such an example, if the agreed-upon lease interval is 2000 milliseconds, the first storage system (714) establishes (708) a synchronous replication lease that is valid until the monotonically increasing clock for the storage controller of the first storage system (714) reaches a value of 3000 milliseconds. If the storage controller of the first storage system (714) does not receive a message from the storage controller of the second storage system (724) containing an updated value of the monotonically increasing clock for the storage controller of the first storage system (714) by the time the monotonically increasing clock for the storage controller of the first storage system (714) reaches a value of 3000 milliseconds, the first storage system (714) may treat the synchronous replication lease as expired and take various actions, which are described in more detail below. The reader will understand that the storage controllers in the remaining storage systems (724, 728) in the pod will react similarly and perform similar tracking and updating of the synchronous replication lease. Essentially, the receiving controller may be guaranteed that the network and its paired controller were running somewhere during that time interval, and that its paired controller received the message it sent somewhere during that time interval. Without any adjustment of the clock, the receiving controller cannot know exactly where in the time interval the network and paired controller were performing, and cannot truly know if there were any queuing delays in sending its clock value or in receiving it back.

[0250] In a pod consisting of two storage systems, each with a simple primary controller, where the primary controllers exchange clocks as part of their cluster communication, each primary controller can use an activity lease to place limits on when it does not know with certainty that its paired controller was running. At the point where the primary controller becomes uncertain (when the controller's active lease on the connection has expired), the primary controller can send a message indicating that it is uncertain and that a properly synchronized connection must be re-established before the active lease can be resumed. If the network is functioning in one direction but not properly in the other, these messages may be received and no response may be received. This may be the first indication by the running paired controller that the connection is not running properly, since its own active lease may not have yet expired due to a different combination of lost messages and queuing delays. As a result, if such a message is received, it should consider its own active lease to have expired as well, and it should begin sending its own messages, attempting to coordinate synchronization of the connection and resumption of the active lease. Until that happens and a new set of clock exchanges is successful, no controller can consider its active lease valid.

[0251] In this model, the controller may wait lease interval seconds after it begins sending a re-establishment message; if the controller does not receive a response, it may be assured that either the paired controller is down or the paired controller's own lease for the connection has expired. To handle small amounts of clock drift, the controller may wait slightly longer than the lease interval (i.e., re-establish the lease). When the controller receives the re-establishment message, it will immediately consider the re-establishment lease to have expired rather than waiting (because the controller knows that sending the controller's active lease has expired), but it often makes sense to try additional messaging before giving up if the message loss was a temporary condition caused, for example, by a congested network switch.

[0252] In an alternative embodiment, in addition to establishing a synchronous replication lease, a cluster membership lease may also be established upon receipt of a clock value from a paired storage system or upon receipt of a clock exchanged with a paired storage system. In such an example, each storage system may have its own synchronous replication lease with every paired storage system and its own cluster membership lease. Expiration of a synchronous replication lease with any pair may result in processing pauses. However, cluster membership cannot be recalculated until cluster membership leases with all pairs have expired. Therefore, the duration of a cluster membership lease should be set based on the interaction of messages and clock values to ensure that a cluster membership lease with a pair does not expire until after the pair's synchronous replication link for that link has expired. The reader will understand that a cluster membership lease may be established by each storage system in a pod and may be associated with a communication link between any two storage systems that are members of the pod. Furthermore, the cluster membership lease may be extended after the expiration of the synchronous replication lease for a period at least as long as the period for the synchronous replication lease expiration. The cluster membership lease may be extended upon receipt of a clock value received from a paired storage system as part of a clock exchange, and the cluster membership lease period from the current clock value may be at least as long as the period established for the last synchronous replication lease extension based on the exchanged clock values. In additional embodiments, additional cluster membership information, including when the session was initially negotiated, may be exchanged over the connection. The reader will understand that in embodiments utilizing cluster membership leases, each storage system (or storage controller) may have its own value for the cluster membership lease. Such a lease should not expire until all synchronous replication leases across all pod members are guaranteed to have expired, given that a cluster lease expiration allows new membership to be established, for example, through a mediator competition, and that a synchronous replication lease expiration forces a pause in the processing of new requests. In such instances, quiescing must be guaranteed to occur everywhere before cluster membership actions can be taken.

[0253] The reader will now be familiar with identifying (702) multiple storage systems (714, 724, 728) across which the dataset (712) is synchronously replicated, with only one of the storage systems (714) being a data set (712); configuring (704) one or more data communication links (716, 718, 720) between each of the multiple storage systems (714, 724, 728) used to synchronously replicate the dataset (712); and configuring the multiple storage systems (714, 724) to synchronously replicate the dataset (712). While the illustration shows exchanging (706) timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728) between the plurality of storage systems (714, 724, 728) and establishing (708) a synchronous replication lease in accordance with the timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728), it is understood that the remaining storage systems (724, 728) may also perform such steps. In practice, establishing a synchronous replication relationship between two or more storage systems (714, 724, 728) may require cooperation and interaction between two or more storage systems (714, 724, 728), so all three storage systems (714, 724, 728) may simultaneously perform one or more of the steps described above.

For further explanation, FIG. 8 sets forth a flowchart illustrating an additional example method for establishing a synchronous replication relationship between two or more storage systems (714, 724, 728) according to some embodiments of the present disclosure. The example method illustrated in FIG. 8 includes identifying (702), for a data set (712), a plurality of storage systems (714, 724, 728) across which the data set (712) is to be synchronously replicated; configuring (704) one or more data communication links (716, 718, 720) between each of the plurality of storage systems (714, 724, 728) used to synchronously replicate the data set (712); and configuring (705) one or more data communication links (716, 718, 720) between the plurality of storage systems (714, 724, 728) to synchronously replicate the data set (712). The example method illustrated in FIG. 8 is similar to the example method illustrated in FIG. 46 because it also includes exchanging (706) timing information (710, 722, 726) for at least one of the storage systems (714, 724, 728) and establishing a synchronous replication lease in accordance with the timing information (710, 722, 726) for at least one of the storage systems (714, 724, 728), where the synchronous replication lease identifies a period of time during which the synchronous replication relationship is valid.

8, establishing a synchronous replication lease (708) according to timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728) may include coordinating (802) clocks between the plurality of storage systems (714, 724, 728). In the example method shown in FIG. 8, coordinating (802) clocks between the plurality of storage systems (714, 724, 728) may be performed, for example, by exchanging one or more messages sent between the storage systems (714, 724, 728). The one or more messages sent between the storage systems (714, 724, 728) may include information, such as a clock value of the storage system whose clock value is used by all other storage systems, instructions to set the clock value for all storage systems to a predetermined value, and a confirmation message from the storage system that its clock value has been updated. In such examples, the storage systems (714, 724, 728) may be configured such that a clock value for a particular storage system (e.g., a leader storage system) should be used by all other storage systems, clock values from all of the storage systems that meet some particular criteria (e.g., highest clock value) should be used by all other storage systems, etc. In such examples, some predetermined amount of time may be added to a clock value received from another storage system to account for transmission time associated with the exchange of messages.

8, establishing (708) a synchronous replication lease in accordance with timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728) may include exchanging (804) uncoordinated clocks among the plurality of storage systems (714, 724, 728). Exchanging (804) uncoordinated clocks among the plurality of storage systems (714, 724, 728) may be performed, for example, by the storage controllers of each storage system (714, 724, 728) exchanging values of local monotonically increasing clocks, as described in more detail above. In such an example, each storage system (714, 724, 728) may utilize agreed-upon synchronous replication lease intervals and messaging received from the other storage systems (714, 724, 728) to establish (708) a synchronous replication lease.

[0257] The example method shown in FIG. 8 also includes postponing (806) processing of I/O requests received after a synchronous replication lease has expired. I/O requests received by any of the storage systems after a synchronous replication lease has expired may be postponed for a predetermined amount of time sufficient to attempt to re-establish the synchronous replication relationship, e.g., until a new synchronous replication lease is established. In such an example, the storage system may postpone processing of the I/O request by failing with some type of "busy" or temporary failure indication, or in some other manner (806).

For further explanation, FIG. 9 sets forth a flowchart illustrating an additional example method for establishing a synchronous replication relationship between two or more storage systems (714, 724, 728) according to some embodiments of the present disclosure. The example method illustrated in FIG. 9 includes identifying (702), for a data set (712), a plurality of storage systems (714, 724, 728) across which the data set (712) is to be synchronously replicated; configuring (704) one or more data communication links (716a, 716b, 718a, 718b, 720a, 720b) between each of the plurality of storage systems (714, 724, 728) used to synchronously replicate the data set (712); and configuring (705) one or more data communication links (716a, 716b, 718a, 718b, 720a, 720b) between the plurality of storage systems (714, 724, 728). The example method illustrated in FIG. 9 is similar to the example method illustrated in FIG. 46 because it also includes exchanging (706) timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728) and establishing (708) a synchronous replication lease in accordance with the timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728), where the synchronous replication lease identifies a period of time during which the synchronous replication relationship is valid.

9, configuring (704) one or more data communication links (716a, 716b, 718a, 718b, 720a, 720b) between each of the plurality of storage systems (714, 724, 728) used to synchronously replicate the dataset (712) may include configuring (902) a data communication link (716a, 716b, 718a, 718b, 720a, 720b) between each of the plurality of storage systems (714, 724, 728) used to synchronously replicate the dataset (712) for each of a plurality of data communication types. In the example method shown in FIG. 9, each storage system may be configured to generate a plurality of data communication types that the storage system transmits to other storage systems in the pod. For example, a storage system may generate a first type of data communication that includes data that is part of an I/O operation (e.g., data written to the storage system as part of a write request issued by a host), may be configured to generate a second type of data communication that includes configuration changes (e.g., information generated in response to creating, expanding, deleting, or renaming a volume), and may be configured to generate a third type of data communication that includes communications related to detecting and handling storage system and interconnect failures. In such examples, the data communication type may be determined, for example, based on which software module initiated the message, based on which hardware component initiated the message, based on the type of event that initiated the message, and in other manners. In the example method shown in FIG. 9, configuring (902) a data communication link (716a, 716b, 718a, 718b, 720a, 720b) between each of the plurality of storage systems (714, 724, 728) for each of the plurality of data communication types may be performed, for example, by configuring the storage systems to use a separate interconnect for each of the plurality of data communication types, by configuring the storage systems to use a separate network for each of the plurality of data communication types, or in other ways.

9 also includes detecting (904) that a synchronous replication lease has expired. In the example method shown in FIG. 9, detecting (904) that a synchronous replication lease has expired may be performed, for example, by a particular storage system comparing its current clock value to the period during which the lease was valid. The storage systems (714, 724, 728) are configured to adjust their clocks to set the value of the clock within each storage system (714, 724, 728) to a value of 5000 milliseconds, and each storage system (714, 724, 728) establishes (708) a synchronous replication lease that extends beyond that clock value for a lease interval of 2000 milliseconds, whereby the synchronous replication lease for each storage system (714, 724, 728) has expired when the clock within the particular storage system (714, 724, 728) reaches a value greater than 7000 milliseconds. In such an example, detecting that a synchronous replication lease has expired (904) may be performed by determining that a clock within a particular storage system (714, 724, 728) has reached a value of 7001 milliseconds or greater.

[0261] The reader will understand that the occurrence of other events may also cause each storage system (714, 724, 728) to immediately treat the synchronous replication lease as expired. For example, a storage system (714, 724, 728) may immediately treat the synchronous replication lease as expired upon detecting a communication failure between the storage system (714, 724, 728) and another storage system (714, 724, 728) in the pod; a storage system (714, 724, 728) may immediately treat the synchronous replication lease as expired upon receiving a lease re-establishment message from another storage system (714, 724, 728) in the pod; a storage system (714, 724, 728) may immediately treat the synchronous replication lease as expired upon detecting that another storage system (714, 724, 728) in the pod has failed; and so on. In such an example, the occurrence of any of the events described in the preceding paragraphs may cause the storage system to detect (904) that the synchronous replication lease has expired.

9 also includes reestablishing the synchronous replication relationship (906). In the example method shown in FIG. 9, reestablishing the synchronous replication relationship (906) may be performed, for example, through the use of one or more reestablishment messages. Such reestablishment messages may include, for example, an identification of the pod with which the synchronous replication relationship is to be reestablished, information needed to configure one or more data communication links, updated timing information, etc. In this manner, the storage systems (714, 724, 728) may each identify (702) for a data set (712) a plurality of storage systems (714, 724, 728) across which the data set (712) is to be synchronously replicated; configure (704) one or more data communication links (716a, 716b, 718a, 718b, 720a, 720b) between each of the plurality of storage systems (714, 724, 728) used to synchronously replicate the data set (712); and configure (705) one or more data communication links (716a, 716b, 718a, 718b, 720a, 720b) between the plurality of storage systems (714, 724, 728). The synchronous replication relationship may be re-established (906) in substantially the same manner as the synchronous replication relationship was originally created, including, but not limited to, performing one or more of: exchanging (706) timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728); and establishing (708) a synchronous replication lease in accordance with the timing information (710, 722, 726) for at least one of the plurality of storage systems (714, 724, 728), the synchronous replication lease identifying a period of time during which the synchronous replication relationship is valid.

9, expiration of a synchronous replication lease may be followed by some set of events, such as a re-establishment message followed by a new activity event or some other action. Data communications, configuration communications, or other communications may be in transition while the synchronous replication lease expires and is re-established. In fact, communications may not be received until, for example, after a new synchronous replication lease is established. In such cases, communications may have been sent based on one understanding of the state of a pod, cluster, or network link, and may now be received by a storage system (714, 724, 728) that has a different understanding of one aspect or another of that state. Thus, generally, there should be some means of ensuring that communications received are discarded if they were sent before some set of clusters or link states change. One way to ensure that communications sent before any convergence of the cluster or received communications are discarded if link conditions change is to establish some session identifier (e.g., a number) associated with establishing or re-establishing the link with an extended working synchronous replication lease. After cluster communications are re-established, the link gets a new session identifier. This identifier may be included with data messages, configuration messages, or other communication messages. Any message received with an incorrect session identifier is discarded or results in an error response indicating a session identifier mismatch.

[0264] The reader will understand that how a storage system (714, 724, 728) responds to the re-establishment of a synchronous replication lease may vary based on the different implementations of the storage system and pod. In the case of a simple primary controller with two storage systems, a new request to perform an operation (read, write, file operation, object operation, management operation, etc.) on the storage system that is received after the controller's synchronous replication lease has expired may have its processing postponed, aborted, or fail with some kind of "retry later" error code. Thus, a running primary storage controller may be guaranteed not to process new requests if its paired storage controller's synchronous replication lease has expired, and it may be certain when its own re-established lease has expired. After the re-established lease has expired, it is safe for the controller to consider its paired controller offline and then consider corrective actions, including continuing storage processing without the paired controller. Exactly what those actions may be may vary based on a wide variety of considerations and implementation details.

[0265] In the case of storage systems with primary and secondary controllers, the primary controller still running on a storage system may attempt to connect to the former second controller of the paired storage system, assuming that the former second controller of the paired storage system may take over. Alternatively, the primary controller still running on a storage system may wait a certain amount of time, perhaps a maximum secondary takeover time. If the secondary controller connects and establishes a new connection with a new synchronous replication lease within a reasonable amount of time, the pod may then recover itself to a consistent state (described below) and then proceed normally. If the paired secondary controller does not connect quickly enough, then the still-running primary controller may take additional action, such as, for example, considering the paired storage system to have failed and then attempting to determine whether it should continue operating without the paired storage system. The primary controller may instead maintain an active lease connection to a secondary controller on the paired storage system in the pod. In that case, expiration of the primary-to-primary re-established lease may instead result in the surviving primary using that connection to query the secondary for takeover, rather than first requiring that the connection be established. Two primary storage controllers may be running while the network is not functioning between them, but the network is functioning between one or the other primary controller and the paired secondary controller. In that case, high availability monitoring within the storage system may not by itself detect the condition that triggers a failover from the primary to the secondary controller. Responses to that condition may include triggering a primary-to-secondary failover anyway to simply resume synchronous replication, routing communication traffic from the primary through the secondary, or acting exactly as if communication had completely failed between the two storage systems, resulting in the same failure processing that occurred.

[0266] When multiple controllers are active for a pod (including both dual active-active controller storage systems and scale-out storage systems), leases may still be maintained through individual controller cluster communication with any or all controllers in the paired storage systems. In this case, an expired synchronous replication lease may lead to a stall in new request processing for the pod across the storage system. The leasing model may be extended with an exchange of clocks and paired clock responses between all active controllers in the storage system, with the additional exchange of their clocks with every paired controller in the paired storage system. If there is an operational path in which a particular local controller's clock is exchanged with any paired controller, the controllers can then use that path for independent synchronous replication leases and possibly for independent re-establishment of leases. In this case, local controllers in a storage system may further exchange clocks among each other for local leases between each other as well. This may already be incorporated into the local storage system's high availability and monitoring mechanisms, but any timing related to the storage system's high availability mechanisms should be taken into account during the duration of the active and re-established leases, or any additional delays between the re-established lease expiration and the action taken to handle the interconnect failure.

[0267] Alternatively, storage system-to-storage system cluster communication or leasing protocols alone may be assigned to one primary controller at a time in each multi-controller or scale-out system, at least for a particular pod. This service may migrate from controller to controller as a result of failure, or perhaps as a result of load imbalance. Alternatively, cluster communication or leasing protocols may run on a subset of controllers (e.g., two) to limit the complexity of clock exchange or analyzing failure conditions. Each local controller may need to exchange clocks among controllers that process storage system-to-storage system leases, and the time to respond after lease expiration must be adjusted accordingly, to account for potential cascading delays when individual controllers can be sure that they have achieved processing pauses. Connections that are not currently relied upon for leases related to processing pauses may still be monitored for alerts.

9 also includes attempting to take over I/O processing for the dataset (908). In the example method shown in FIG. 9, attempting to take over I/O processing for dataset (712) (908) may be performed by storage systems (714, 724, 728), for example, racing to a mediator. If a particular storage system (714, 724, 728) successfully takes over I/O processing for dataset (712), all accesses of dataset (712) are serviced by the particular storage system (714, 724, 728) until a synchronous replication relationship can be reestablished, and any changes to dataset (712) that occurred after the previous synchronous replication relationship expired may then be transferred and persisted on the other storage system (714, 724, 728). In such an example, an attempt 908 to take over I/O processing for dataset 712 may occur only after the expiration of some period of time after the synchronous replication lease expires. For example, an attempt to determine how to proceed after a link failure (including one or more of the storage systems attempting to take over I/O processing for the dataset) may not begin until a period of time after the synchronous replication lease expires, i.e., at least as long as the maximum lease time resulting from a clock exchange.

[0269] The reader will understand that while in many of the examples shown above, only one of the storage systems (714) is shown performing the steps described above, establishing a synchronous replication relationship between two or more storage systems may require cooperation and interaction between two or more storage systems, and therefore, in practice, all of the storage systems (714, 724, 728) in a pod (or in a pod being formed) may simultaneously perform one or more of the steps described above.

[0270] For further explanation, FIG. 10 sets forth a flowchart illustrating an additional example method for establishing a synchronous replication relationship between two or more storage systems (1024, 1046) according to some embodiments of the present disclosure. While the example method illustrated in FIG. 10 illustrates an embodiment in which the dataset (1022) is synchronously replicated across only two storage systems (1024, 1046), the example illustrated in FIG. 10 may be extended to embodiments in which the dataset (1022) is synchronously replicated across additional storage systems that may perform steps similar to those performed by the two illustrated storage systems (1024, 1046).

10 includes configuring (1002), by storage system (1024), one or more data communication links (1052) between storage system (1024, 1046) and a second storage system (1046). In the example method shown in FIG. 10, storage system (1024) may configure (1002) one or more data communication links (1052) between storage system (1024) and the second storage system (1046), for example, by identifying a port defined over a data communication network used to exchange data communication with the second storage system (1046), by identifying a point-to-point data communication link used to exchange data communication with the second storage system (1046), or in various other ways. If secure communication is required, some form of key exchange may be required, or the communication may occur or be initiated through some service, such as Secure Shell (SSH), SSL, or some other service or protocol built around public key or Diffie-Hellman key sharing, or a suitable alternative. Secure communication may also be mediated through some vendor-provided cloud service that is tied in some way to a customer identity. Alternatively, a service configured to run at the customer premises, e.g., running in a virtual machine or virtual container, could be used to broker the key exchange necessary for secure communication between the replicating storage systems (1024, 1046). The reader will understand that a pod containing more than two storage systems may require communication links between most or all of the individual storage systems. In the example method shown in FIG. 10, the second storage system (1046) may similarly configure one or more data communication links (1052) between the storage system (1024) and the second storage system (1046) (1026).

10 also includes transmitting (1004) timing information (1048) for storage system (1024) from storage system (1024) to second storage system (1046). The timing information (1048) for storage system (1024) may be embodied, for example, as the value of a clock within storage system (1024), as a representation of a clock value (e.g., a sequence number that storage system (1024) can first record), as the most recently received value of a clock within second storage system (1046), etc. In the example method shown in FIG. 10 , storage system (1024) may transmit timing information (1048) for storage system (1024) to second storage system (1046) (1004), for example, via one or more messages transmitted from storage system (1024) to second storage system (1046) over a data communication link (1052) between the two storage systems (1024, 1046). In the example shown in FIG. 10 , second storage system (1046) may similarly transmit timing information (1050) for second storage system (1046) from second storage system (1046) to storage system (1024) (1030).

10, transmitting (1004) timing information (1048) for storage system (1024) from storage system (1024) to second storage system (1046) may include transmitting (1006) the value of a clock for storage system (1024). In the example shown in FIG. 10, storage system (1024) may transmit (1006) the value of its clock to second storage system (1046) as part of an effort to coordinate the clocks between storage systems (1024, 1046). In such an example, storage system (1024) may include a local monotonically increasing clock whose value is transmitted (1006) via one or more messages sent to second storage system (1046) via data communication link (1052) between the two storage systems (1024, 1046). In the example method shown in FIG. 10, transmitting (1030) timing information (1050) for the second storage system (1046) from the second storage system (1046) to the storage system (1024) may also include transmitting (1032) the value of a clock of the second storage system (1046).

10, transmitting (1004) timing information (1048) for storage system (1024) from storage system (1024) to second storage system (1046) may also include transmitting (1008) the most recently received value of the clock of the second storage system (1046). In the example method shown in FIG. 10, transmitting (1008) the most recently received value of the clock of the second storage system (1046) may be performed, for example, as part of an effort to eliminate the need to adjust clocks between storage systems (1024, 1046) while still achieving timing guarantees. In such an embodiment, each storage system (1024, 1046) may have a local monotonically increasing clock. A synchronous replication lease may be established between storage systems (1024, 1048) by each storage system (1024, 1048) sending its clock value to the other storage system (1024, 1048) along with the last clock value it received from the other storage system (1024, 1048). When a particular storage system (1024, 1048) also receives a clock value from another storage system (1024, 1048), the particular storage system (1024, 1048) may add any agreed-upon lease interval to the received clock value and use it to establish a synchronous replication lease. In the example method shown in FIG. 10, transmitting (1030) timing information (1050) for the second storage system (1046) from the second storage system (1046) to the storage system (1024) may also include transmitting (1034) the most recently received value of the clock of the storage system (1024).

10 also includes receiving (1010) timing information (1050) for the second storage system (1046) by the storage system (1024) from the second storage system (1046). In the example method shown in FIG. 10, the storage system (1024) may receive (1010) the timing information (1050) for the second storage system (1046) from the second storage system (1046) via one or more messages transmitted from the second storage system (1046) via a data communication link (1052) between the two storage arrays (1024, 1046). In the example shown in FIG. 10, the second storage system (1046) may similarly receive (1028) the timing information for the storage system (1024) from the storage system (1024).

10 also includes setting (1012) a clock value at storage system (1024) according to timing information (1050) for second storage system (1046). In the example shown in FIG. 10, setting (1012) a clock value at storage system (1024) according to timing information (1050) for second storage system (1046) may be performed, for example, as part of an effort to coordinate clocks between the two storage systems (1024, 1046). In such an example, the two storage systems (1024, 1046) may be configured, for example, to set their respective clock values to a value that is some predetermined amount higher than the highest clock value between the storage systems (1024, 1046), to set their respective clock values to a value equal to the highest clock value between the pair of storage systems (1024, 1046), to set their respective clock values to a value generated by applying some function to the respective clock values of each storage system (1024, 1046), or in some other manner. In the example method shown in FIG. 10 , the second storage system (1046) may similarly set (1036) its clock value in accordance with the timing information (1048) for storage system (1024).

10 also includes establishing (1014) a synchronous replication lease. In the example method shown in FIG. 10, establishing (1014) a synchronous replication lease may be implemented, for example, by establishing a synchronous replication lease that extends for some predetermined lease interval beyond the equivalent clock values between the two storage systems (1024, 1046), by establishing a synchronous replication lease that extends for some predetermined lease interval beyond the unadjusted clock value associated with one of the storage systems (1024, 1046), or in some other manner. In the example method shown in FIG. 19, the second storage system (1046) may similarly set (1036) its clock value in accordance with the timing information (1048) for storage system (1024).

10 also includes detecting (1016) by the storage system (1024) that the synchronous replication lease has expired. In the example method shown in FIG. 10, detecting (1016) that the synchronous replication lease has expired may be performed, for example, by the storage system (1024) comparing the current clock value to the period during which the lease was valid. Consider an example in which the storage systems (1024, 1046) are configured to adjust the value of the clock within each storage system (1024, 1046) to a value of 5000 milliseconds, and each storage system (1024, 1046) establishes (1038) an extended synchronous replication lease for a lease interval of 2000 milliseconds beyond that clock value, thereby causing the synchronous replication lease for each storage system (1024, 1046) to expire when the clock within a particular storage system (1024, 1046) reaches a value greater than 10,000 milliseconds. In such an example, detecting that the synchronous replication lease has expired (1006) may be performed by determining that a clock within the storage system (1024) has reached a value of 10001 milliseconds or greater. In the example method shown in FIG. 10, the second storage system (1046) may similarly detect that the synchronous replication lease has expired (1040).

10 also includes attempting (1020) by storage system (1024) to take over I/O processing for dataset (1022). In the example method shown in FIG. 10, attempting (1020) to take over I/O processing for dataset (1022) may be performed by storage system (1024) dispatching to a mediator, for example. If storage system (1024) successfully takes over I/O processing for dataset (1022), all accesses of dataset (1022) may be serviced by storage system (1024) until the synchronous replication relationship is reestablished, and any changes to dataset (1022) that occurred after the previous synchronous replication relationship expired may be transferred to and persisted in the second storage system (1046). In the example method shown in FIG. 10, the second storage system (1046) may also attempt to take over I/O processing for the dataset (1022) (1044).

[0280] The example method illustrated in FIG. 10 also includes, by the storage system (1024), attempting (1018) to reestablish the synchronous replication relationship. In the example method illustrated in FIG. 10, attempting (1018) to reestablish the synchronous replication relationship may be performed, for example, by using one or more reestablishment messages. Such reestablishment messages may include, for example, identification of the pod with which the synchronous replication relationship is to be reestablished, information needed to configure one or more data communication links, updated timing information, etc. In this manner, the storage system (1024) may reestablish the synchronous replication relationship substantially in the same manner as the synchronous replication relationship was originally created. In the example illustrated in FIG. 10, the second storage system (1046) may similarly attempt (1042) to reestablish the synchronous replication relationship.

[0281] For further explanation, FIG. 11 sets forth a flowchart illustrating an example method for servicing I/O operations directed to a dataset (1142) synchronized across multiple storage systems (1138, 1140) in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (1138, 1140) illustrated in FIG. 11 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 11 may include the same components as the storage systems described above, fewer components, or additional components.

11 may be embodied, for example, as the contents of a particular volume, as the contents of a particular share of a volume, or as any other collection of one or more data elements. The dataset (1142) may be synchronized across multiple storage systems (1138, 1140) such that each storage system (1138, 1140) maintains a local copy of the dataset. In the examples described herein, such dataset (1142) is synchronously replicated across storage systems (1138, 1140) such that the dataset (1142) can be accessed through any of the storage systems (1138, 1140) having performance characteristics such that any one storage system in the cluster operates substantially less optimally than any other storage system in the cluster, at least as long as the cluster and the particular storage system being accessed are nominally performing. In such a system, modifications to a dataset (1142) should be made to copies of the dataset resident on each storage system (1138, 1140) so that accessing the dataset (1142) on any storage system (1138, 1140) produces consistent results. For example, a write request to a dataset must be serviced on all storage systems (1138, 1140), or on none of the storage systems (1138, 1140) that were nominally running at the beginning of the write and remained nominally running until the write was completed. Similarly, some groups of operations (e.g., two write operations directed to the same location in a dataset) must be performed in the same order, or other steps must be taken on all storage systems (1138, 1140), so that the dataset is ultimately identical on all storage systems (1138, 1140). Modifications to the dataset (1142) need not occur at exactly the same time, although some actions (e.g., issuing an acknowledgment that a write request is directed to the dataset and allowing read access to locations within the dataset targeted by the write request that have not yet completed on both storage systems) may be deferred until the copies of the dataset on each storage system (1138, 1140) have been modified.

11, designating one storage system (1140) as a "leader" and another storage system (1138) as a "follower" may refer to the respective relationship of each storage system for purposes of synchronously replicating a particular data set across the storage systems. In such an example, and as described in more detail below, the leader storage system (1140) may be responsible for performing some processing of incoming I/O operations, passing related information to the follower storage system (1138), or performing other tasks not solicited by the follower storage system (1140). The leader storage system (1140) may be responsible for performing tasks not solicited by the follower storage system (1138) for all incoming I/O operations, or alternatively, the leader-follower relationship may be specific to only a subset of the I/O operations received by either storage system. For example, a leader-follower relationship may be specific to I/O operations directed to a first volume, a first group of volumes, a first group of logical addresses, a first group of physical addresses, or some other logical or physical delineator. In this manner, a first storage system may act as a leader storage system for I/O operations directed to a first set of volumes (or other delineators), while a second storage system may act as a leader storage system for I/O operations directed to a second set of volumes (or other delineators). The example method shown in FIG. 11 illustrates an embodiment in which synchronizing the multiple storage systems (1138, 1140) occurs in response to receiving a request (1104) to modify a data set (1142) by a leader storage system (1140), although synchronizing the multiple storage systems (1138, 1140) may also be performed in response to receiving a request (1104) to modify a data set (1142) by a follower storage system (1138), as described in more detail below.

11 includes receiving (1106) a request (1104) by a leader storage system (1140) to modify a dataset (1142). The request (1104) to modify a dataset (1142) may be implemented, for example, as a request to write data to a location in the storage system (1140) containing data included in the dataset (1142), as a request to write data to a volume containing data included in the dataset (1142), as a request to take a snapshot of the dataset (1142), essentially as a virtual range copy, as an UNMAP operation representing the deletion of some portion of the data in the dataset (1142), as a modifying transformation of the dataset (1142) (rather than a change to a portion of the data within the dataset), or as some other operation that causes a change to some portion of the data included in the dataset (1142). In the example method shown in FIG. 11, a request (1104) to modify a dataset (1142) is issued by a host (1102), which may be embodied, for example, as an application running on a virtual machine, as an application running on a computing device connected to the storage system (1140), or as some other entity configured to access the storage system (1140).

11 also includes generating (1108), by the leader storage system (1140), information (1110) describing modifications to the dataset (1142). The leader storage system (1140) may determine the appropriate outcome of overlapping modifications (e.g., the appropriate outcome of two requests modifying the same storage location), for example, by determining ordering versus any other operations that are in progress, and may generate (1108) the information (1110) describing modifications to the dataset (1142), for example, by computing any distributed state changes to common elements of metadata across all members of a pod (e.g., all storage systems to which the dataset is synchronously replicated). The information (1110) describing modifications to the dataset (1142) may be implemented, for example, as system-level information used to describe I/O operations performed on the storage system. The leader storage system (1140) may generate (1108) information describing modifications to the dataset (1142) by processing the request (1104) to modify the dataset (1142) sufficiently to understand what must occur in order to service the request (1104) to modify the dataset (1142). For example, the leader storage system (1140) may determine whether some ordering of execution of the request (1104) to modify the dataset (1142) relative to other requests to modify the dataset (1142) is required, or some other step must be taken to produce equivalent results at each storage system (1138, 1140), as described in more detail below.

[0286] Consider an example in which request (1104) to modify dataset (1142) is implemented as a request to copy blocks from a first address range within dataset (1142) to a second address range within dataset (1142). In such an example, assume that three other write operations (Write A, Write B, Write C) are directed to the first address range of dataset (1142). In such an example, if the leader storage system (1140) services Write A and Write B (but does not service Write C) before copying blocks from a first address range in the dataset (1142) to a second address range in the dataset (1142), then the follower storage system (1138) must also service Write A and Write B (but does not service Write C) before copying blocks from the first address range in the dataset (1142) to the second address range in the dataset (1142) in order to produce consistent results. Thus, when the leader storage system (1140) generates (1108) information (1110) describing modifications to the data set (1142), in this example, the leader storage system (1140) could generate information (e.g., sequence numbers for Write A and Write B) that identifies other operations that must be completed before the follower storage system (1138) can process the request (1104) to modify the data set (1142).

[0287] Consider an additional example in which two requests (e.g., Write A and Write B) are directed to overlapping portions of data set (1142). In such an example, if the follower storage system (1138) services Write B and later services Write A, but the leader storage system (1140) services Write A and later services Write B, the data set (1142) will not be consistent across both storage systems (1138, 1140). Thus, when the leader storage system (1140) generates (1108) information (1110) describing modifications to the data set (1142), in this example, the leader storage system (1140) could generate information identifying the order in which the requests should be executed (e.g., sequence numbers for Write A and Write B). Alternatively, rather than generating information (1110) describing modifications to the dataset (1142) that require intermediate operations from each storage system (1138, 1140), the leader storage system (1140) may generate information (1110) describing modifications to the dataset (1142) that includes information identifying the appropriate outcome of the two requests (1108). For example, if write B logically follows (and overlaps with) write A, the end result should be that dataset (1142) includes portions of write B that overlap with write A, rather than including portions of write A that overlap with write B. Such an outcome would be facilitated by merging the results in memory and writing the results of such a merge to the dataset (1142), rather than strictly requiring a particular storage system (1138, 11410) to perform write A and then subsequently perform write B. The reader will understand that more subtle cases pertain to snapshots and virtual address range copies.

[0288] The reader will further understand that the correct result for any operation must be committed to a point where it is recoverable before receipt of the operation can be acknowledged. However, in some cases, multiple operations can be committed together, and in other cases, operations can be partially committed if recovery ensures correctness. For example, a snapshot could be committed locally with recorded dependencies on expected write A and write B, but A or B itself may not be committed. If receipt of the snapshot cannot be acknowledged and lost I/O cannot be recovered from another array, recovery may end up canceling the snapshot. Also, if write B overlaps with write A, the leader storage system may "instruct" B to be behind A, but A is actually discarded, and then the operation writing A will simply wait for B. Write A, Write B, Write C, and Write D combined with the snapshot between A, B and C, D will be able to commit and/or acknowledge receipt of some or all portions together, as long as recovery does not cause a snapshot inconsistency across the array, and as long as the acknowledgment does not complete a later operation before the earlier operation has persisted to the point where it is guaranteed to be recoverable.

11 also includes transmitting (1112) information (1110) from the leader storage system (1140) to the follower storage system (1138) describing the modifications to the dataset (1142). The transmitting (1112) information (1110) from the leader storage system (1140) to the follower storage system (1138) describing the modifications to the dataset (1142) may be performed, for example, by the leader storage system (1140) sending one or more messages to the follower storage system (1138). The leader storage system (1140) may also transmit an I/O payload (1114) for the request (1104) to modify the dataset (1142) in the same message or in one or more different messages. The I/O payload (1114) may be implemented as data written to storage in the follower storage system (1138), for example, when the request (1104) to modify the dataset (1142) is implemented as a request to write data to the dataset (1142). In such an example, because the request (1104) to modify the dataset (1142) was received (1106) by the leader storage system (1140), the follower storage system (1138) has not yet received the I/O payload (1114) associated with the request (1104) to modify the dataset (1142). In the example shown in FIG. 11 , information (1110) describing modifications to the dataset (1142) and an I/O payload (1142) associated with a request (1104) to modify the dataset (1142) may be transmitted (1112) from the leader storage system (1140) to the follower storage system (1138) via one or more data communications networks coupling the leader storage system (1140) to the follower storage system (1138), via one or more dedicated data communications links (e.g., a first link for transmitting the I/O payload and a second link for transmitting the information describing the modifications to the dataset) coupling the leader storage system (1140) to the follower storage system (1138), or via some other mechanism.

11 also includes receiving (1116) by the follower storage system (1138) information (1110) describing modifications to the dataset (1142). The follower storage system (1138) may receive (1116) the information (1110) describing the modifications to the dataset (1142) and the I/O payload (1114) from the leader storage system (1140), for example, via one or more messages sent from the leader storage system (1140) to the follower storage system (1138). One or more messages may be sent from the leader storage system (1140) to the follower storage system (1138) via one or more dedicated data communication links between the two storage systems (1138, 1140), by the leader storage system (1140) writing the messages to a predetermined storage location (e.g., a queue location) on the follower storage system (1138) using RDMA or a similar mechanism, or in other ways.

[0291] In one embodiment, the follower storage system (1138) may receive information (1110) and I/O payload (1114) describing modifications to the dataset (1142) from the leader storage system (1140) using SCSI requests (write from sender to receiver, or read from receiver to sender) as the communication mechanism (1116). In such an embodiment, the SCSI write request is used to encode the information intended to be sent (including whatever data and metadata) and may be delivered to a special pseudo device, or over a specially configured SCSI network, or through any other agreed-upon addressing mechanism. Alternatively, the model may issue a set of open SCSI read requests from the receiver to the sender using a special device, a specially configured SCSI network, or other agreed-upon mechanism. The encoded information, including the data and metadata, is delivered to the receiver in response to one or more of these open SCSI requests. Such a model can be implemented over a Fibre Channel SCSI network, which is often deployed as a "dark fibre" storage network between data centres. It also enables the use of the same network lines for multipathing from the host to the remote array and for bulk array-to-array communications.

11 also includes processing (1118) the request (1104) to modify the dataset (1142) by the follower storage system (1138). In the example method shown in FIG. 11, the follower storage system (1138) may process (1118) the request (1104) to modify the dataset (1142) by modifying the contents of one or more storage devices (e.g., NVRAM devices, SSDs, HDDs) included in the follower storage system (1138) according to not only the I/O payload (1114) received from the leader storage system (1140) but also the information (1110) describing the modifications to the dataset (1142). Consider an example in which a request (1104) to modify a dataset (1142) is implemented as a write operation directed to a volume included in the dataset (1142), and the information (1110) describing the modification to the dataset (1142) indicates that the write operation can only be performed after a previously issued write operation has been processed. In such an example, processing (1118) the request (1104) to modify the dataset (1142) may be implemented by the follower storage system (1138) first verifying that the previously issued write operation has been processed at the follower storage system (1138), and then writing the I/O payload (1114) associated with the write operation to one or more storage devices included in the follower storage system (1138). In such an example, a request (1104) to modify a dataset (1142) may be considered completed and successfully processed, for example, when the I/O payload (1114) has been committed to persistent storage in the follower storage system (1138).

11 also includes acknowledging (1120) by the follower storage system (1138) to the leader storage system (1140) receipt of completion of the request (1104) to modify the data set (1142). In the example method shown in FIG. 11, acknowledging (1120) by the follower storage system (1138) to the leader storage system (1140) receipt of completion of the request (1104) to modify the data set (1142) may be performed by the follower storage system (1138) sending an acknowledgement (1122) message to the leader storage system (1140). Such a message may include information identifying the particular request (1104) to modify the completed dataset (1142), as well as any additional information useful, for example, in acknowledging (1120) receipt of completion of the request (1104) to modify the dataset (1142) by the follower storage system (1138). In the example method shown in FIG. 11, acknowledging (1120) receipt of completion of the request (1104) to modify the dataset (1142) to the leader storage system (1140) is indicated by the follower storage system (1138) issuing an acknowledgement (1122) message to the leader storage system (1140).

11 also includes processing (1124) the request (1104) to modify the dataset (1142) by the leader storage system (1140). In the example method shown in FIG. 11, the leader storage system (1140) may process (1124) the request (1104) to modify the dataset (1142) by modifying not only the I/O payload (1114) received as part of the request (1104) to modify the dataset (1142), but also the contents of one or more storage devices (e.g., NVRAM devices, SSDs, HDDs) included in the leader storage system (1140) according to information (1110) describing the modifications to the dataset (1142). Consider an example in which a request (1104) to modify a dataset (1142) is implemented as a write operation directed to a volume included in the dataset (1142), and the information (1110) describing the modification to the dataset (1142) indicates that the write operation can only be performed after a previously issued write operation has been processed. In such an example, processing (1124) the request (1104) to modify the dataset (1142) may be implemented by the leader storage system (1140) first verifying that the previously issued write operation has been processed by the leader storage system (1140), and then writing the I/O payload (1114) associated with the write operation to one or more storage devices included in the leader storage system (1140). In such an example, a request (1104) to modify a dataset (1142) may be considered completed and successfully processed, for example, when the I/O payload (1114) has been committed to persistent storage in the reader storage system (1140).

11 also includes receiving (1126) an indication from the follower storage system (1138) that the follower storage system (1138) has processed the request (1104) to modify the dataset (1136). In this example, the indication that the follower storage system (1138) has processed the request (1104) to modify the dataset (1136) is embodied as an acknowledgment (1122) message sent from the follower storage system (1138) to the leader storage system (1140). The reader will understand that although many of the steps described above are shown and described as occurring in a particular order, no particular order is actually required. In practice, the follower storage system (1138) and the leader storage system (1140) are independent storage systems, and therefore each storage system may be performing some of the steps described above in parallel. For example, the follower storage system (1138) may receive (1116) information (1110) describing modifications to the dataset (1142), may process (1118) the request (1104) to modify the dataset (1142), or may confirm (1120) receipt of completion of the request (1104) to modify the dataset (1142) before the leader storage system (1140) processes (1124) the request (1104) to modify the dataset (1142). Alternatively, the leader storage system (1140) may have processed (1124) the request (1104) to modify the dataset (1142), processed (1118) the request (1104) to modify the dataset (1142), or acknowledged (1120) receipt of completion of the request (1104) to modify the dataset (1142) before the follower storage system (1138) received (1116) the information (1110) describing the modifications to the dataset (1142).

11 also includes acknowledging (1134) receipt of completion of the request (1104) to modify the dataset (1142) by the leader storage system (1140). In the example method shown in FIG. 11, acknowledging (1134) receipt of completion of the request (1104) to modify the dataset (1142) may be performed using one or more acknowledgment (1136) messages sent from the leader storage system (1140) to the host (1102) or via some other suitable mechanism. In the example method shown in FIG. 11, the leader storage system (1140) may determine (1128) whether the request (1104) to modify the dataset (1142) has been processed (1118) by the follower storage system (1138) before acknowledging (1134) receipt of completion of the request (1104) to modify the dataset (1142). The leader storage system (1140) may determine (1128) whether the request (1104) to modify the dataset (1142) has been processed (1118) by the follower storage system (1138), for example, by determining whether the leader storage system (1140) has received an acknowledgment message or other message from the follower storage system (1138) indicating that the request (1104) to modify the dataset (1142) has been processed (1118) by the follower storage system (1138). In such an example, if the leader storage system (1140) affirmatively determines (1130) that the request (1104) to modify the dataset (1142) has been processed (1118) by the follower storage system (1138) and has also been processed (1118) by the leader storage system (1138), the leader storage system (1140) may proceed by acknowledging (1134) receipt of completion of the request (1104) to modify the dataset (1142) to the host (1102) that initiated the request (1104) to modify the dataset (1142). However, if the leader storage system (1140) determines that the request (1104) to modify the dataset (1142) has not yet been processed (1118) (1132) by the follower storage system (1138) or processed (1124) by the leader storage system (1138), the leader storage system (1140) may acknowledge receipt of completion of the request (1104) to modify the dataset (1142) (1134) only when the request (1104) to modify the dataset (1142) has been successfully processed on all storage systems (1138, 1140) across which the dataset (1142) is synchronously replicated, and therefore the leader storage system (1140) may not yet acknowledge receipt of completion of the request (1104) to modify the dataset (1142) (1134) from the host (1102) that initiated the request (1104) to modify the dataset (1142).

[0297] The reader will note that in the example method shown in FIG. 11 , the sending (1112) of information (1110) describing modifications to a data set (1142) from a leader storage system (1140) to a follower storage system (1138), and the acknowledgment (1120) by the follower storage system (1138) to the leader storage system (1140) of receipt of completion of a request (1104) to modify the data set (1142) may be accomplished using a single round-trip message. Single round-trip messaging may be used, for example, with Fibre Channel as the data interconnect. Typically, the SCSI protocol is used with Fibre Channel. Because some older replication technologies can be built to essentially replicate data as SCSI transactions over a Fibre Channel network, such interconnects are commonly set up between data centers. Furthermore, traditional Fibre Channel SCSI infrastructures have had less overhead and lower latency than Ethernet and TCP/IP-based networks. Additionally, when data centers use Fibre Channel and are internally connected to disconnect storage arrays, the Fibre Channel network may be extended to other data centers so that hosts in one data center can switch to accessing a storage array at a remote data center if the local storage array fails.

[0298] SCSI could be used as a general communications mechanism, even though it is typically designed for use with block storage protocols for storing and retrieving data on block-oriented volumes (or for tape). For example, SCSI READ or SCSI WRITE could be used to send or retrieve message data between storage controllers in paired storage systems. A typical implementation of a SCSI WRITE requires two message round-trips: the SCSI initiator sends a SCSI CDB describing the SCSI WRITE operation, the SCSI target receives the CDB, and the SCSI target sends a "ready to receive" message to the SCSI initiator. The SCSI initiator then sends the data to the SCSI target, and once the SCSI WRITE is complete, the SCSI target responds with a successful completion to the SCSI initiator. A SCSI READ request, on the other hand, requires only one round-trip. That is, a SCSI initiator sends a SCSI CDB describing a SCSI READ operation, and the SCSI target receives the CDB and responds with the data and then a successful completion. As a result, in terms of distance, a SCSI READ incurs half the distance-related latency as a SCSI WRITE. For this reason, it may be faster for a data communication receiver to receive a message using a SCSI READ request than for the message sender to send data using a SCSI WRITE request. Using SCSI READ simply requires that the message sender act as a SCSI target and the message receiver act as a SCSI initiator. A message receiver may send several SCSI CDB READ requests to any message sender, and the message sender will respond to one of the pending CDB READ requests when message data is available. The SCSI subsystem may time out a READ request if it is pending for too long (e.g., 10 seconds), so a READ request should be responded to within a few seconds, even if there is no message to send.

[0299] SCSI tape requests support variable response data, which may be more flexible for returning variable-sized message data, as described in the SCSI Stream Commands standard of the T10 Technical Committee of the International Committee on Information Technology Standards. The SCSI standard also supports an immediate mode for SCSI WRITE requests, which could allow for a single round-trip SCSI WRITE command. The reader will understand that many of the embodiments described below also utilize single-round-trip messaging.

12 illustrates a flowchart showing an example method for servicing I/O operations directed to a data set (1142) synchronized across multiple storage systems (1138, 1140, 1150) in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (1138, 1140, 1150) shown in FIG. 11 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system shown in FIG. 11 may include the same components as the storage systems described above, fewer components, or additional components. The example method shown in FIG. 12 is similar to the example method shown in FIG. 11. The example method shown in FIG. 12 also includes receiving (1106) a request to modify (1142) a data set (1142) by a leader storage system (1140), generating (1108) information (1110) by the storage system (1140) describing the modifications to the data set (1142), transmitting (1112) the information (1110) describing the modifications to the data set (1142) from the leader storage system (1140) to a follower storage system (1138), receiving (1116) the information (1110) describing the modifications to the data set (1142) by the follower storage system (1138), and transmitting (1116) the information (1110) describing the modifications to the data set (1142) to the follower storage system (1138). The method is similar to the example method shown in FIG. 11 , as it also includes processing (1118) the request (1104) to modify the dataset (1142) by the follower storage system (1138), acknowledging (1120) receipt of completion of the request (1104) to modify the dataset (1142) to the leader storage system (1140) by the follower storage system (1138), processing (1124) the request (1104) to modify the dataset (1142) by the leader storage system (1140), and acknowledging (1134) receipt of completion of the request (1104) to modify the dataset (1142) by the leader storage system (1140).

[0301] However, the example method shown in FIG. 12 differs from the example method shown in FIG. 11 because it illustrates an embodiment in which a dataset (1142) is synchronously replicated across three storage systems, one of which is a leader storage system (1140) and the remaining storage systems are follower storage systems (1138, 1150). In such an example, the additional follower storage system (1150) performs many of the same steps as the follower storage system (1138) shown in FIG. 11, such as receiving (1142) information (1110) from the leader storage system (1140) describing modifications to the data set (1142), processing (1142) requests (1104) to modify the data set (1142) according to the information (1110) describing the modifications to the data set (1142), and confirming (1146) receipt of completion of the request (1104) to modify the data set (1142) to the leader storage system (1140), such as by using an acknowledgement (1148) message or other suitable mechanism.

[0302] In the example method shown in FIG. 12, the information (1110) describing modifications to a dataset (1142) may include ordering information (1152) for a request (1104) to modify the dataset (1142). In the example method shown in FIG. 12, the ordering information (1152) for a request (1104) to modify the dataset (1142) may express a description of the relationship between the operation (e.g., a request to modify the dataset) and common metadata updates that may be described by the reader storage system (1140) as a set of interdependencies between separate requests to modify the dataset, and possibly between requests to modify the dataset and various metadata changes. These interdependencies may be described as a set of precursors on which a request to modify the dataset depends in some way, and as predicates that must be true for the request to modify the dataset to be completed.

[0303] A queue predicate is an example of a predicate that must be true for a request to modify a dataset to complete. A queue predicate may specify that a particular request to modify a dataset may not complete until a previous request to modify the dataset has completed. A queue predicate may be used, for example, for overlapping write type operations. In such an example, the reader storage system (1140) may declare that a second write type operation logically follows a first such operation, such that the second write type operation cannot complete until the first write type operation completes. Depending on the implementation, the second write type operation may not be made durable until the first such write type operation is assured to be durable (the two operations may both be made durable). A queuing predicate can also be used for snapshot operations and virtual block range copy operations by declaring that a known set of incomplete precursor operations (e.g., a set of write-type operations) must each complete before the snapshot can complete; as additional operations are identified as following the snapshot (before the snapshot completes), each of these operations may presuppose that the snapshot operation itself completes. The predicate could also indicate that these subsequent operations apply to the post-snapshot image of the volume rather than being included in the snapshot.

[0304] An alternative predicate that could be used for snapshots is to assign an identifier to every snapshot and associate all modification operations that can be included in a particular snapshot with that identifier. As a result, a snapshot can be completed when all of the modification operations it contains are complete. This can be done with a count predicate. Each storage system across which a dataset is synchronously replicated can perform its own count of operations associated with time since the last snapshot or some other infrequent operation (alternatively, for embodiments implementing multiple leader storage systems, a count can be established by a particular leader storage system for the portion of the dataset it controls, with those operations orchestrated by that leader storage system). The snapshot operation itself may then include a count predicate that depends on the number of operations that are received and made persistent before the snapshot itself can be made persistent or signaled as completed. Modification operations that should follow the snapshot (before the snapshot is complete) may either be deferred given a queue predicate that depends on the snapshot, or the snapshot identity may be used as an indication that modification operations should be excluded from the snapshot. A virtual block range copy (SCSI EXTENDED COPY or similar operation) may use a queue predicate, or it may use a count predicate and a snapshot or similar identifier. With a count predicate and a snapshot or virtual copy identifier, each virtual block range copy may establish a new virtual snapshot or virtual copy identifier, even if the copy operation covers only two small regions of one or two volumes. In the example described above, the request (1104) to modify dataset (1142) may include a request to take a snapshot of dataset (1142), and therefore the ordering information (1152) for the request (1104) to modify dataset (1142) may include the identification of one or more other requests to modify the dataset that must be completed before taking a snapshot of dataset (1142).

12, the information (1110) describing modifications to the dataset (1142) may include common metadata information (1154) associated with the request (1104) to modify the dataset (1142). The common metadata information (1154) associated with the request (1104) to modify the dataset (1142) may be used to ensure common metadata associated with the dataset (1142) within the storage systems (1138, 1140, 1150) across which the dataset (1142) is synchronously replicated. This associated common metadata may be embodied as any data other than content stored in the dataset (1142), for example, by one or more requests (e.g., one or more write requests issued by a host). Common metadata may include data that the synchronous replication implementation maintains consistency across the storage systems (1138, 1140, 1150) across which the dataset (1142) is synchronously replicated, particularly if the common metadata relates to how the stored content is managed, recovered, resynchronized, snapshotted, or synchronously replicated. The reader will understand that two or more modification operations may depend on the same common metadata; ordering of the modification operations themselves is unnecessary, but consistent application of the common metadata once rather than twice is required. One way to handle multiple dependencies on common metadata is to define the metadata in separate operations that are instantiated and described from the reader storage system. Two modification operations that depend on that common metadata may then be given queue predicates that depend on the modification operations. Another way to handle multiple dependencies on common metadata is to associate the common metadata with the first of two operations and make the second operation dependent on the first. A variation is to have the second operation depend only on the common metadata aspect of the first operation, so that only that part of the first operation must be made persistent before the second operation can be processed. Yet another way to handle multiple dependencies on common metadata is to include the common metadata in all operation descriptions that depend on it. This works well when applying the common metadata is power-based, for example, by simply attaching an identifier to the common metadata. If the identifier has already been processed, it can be ignored. In some cases, the identifier may be associated with a portion of the common metadata.

12, receiving (1126) an indication that the follower storage system has processed (1104) the request to modify the data set (1142) may include receiving (1156) an indication from each of the follower storage systems (1138, 1150) that the follower storage system (1138, 1150) has processed (1104) the request to modify the data set (1142). In this example, the indication that each follower storage system (1138, 1150) has processed (1104) the request to modify the data set (1136) is embodied as a separate acknowledgement (1122, 1148) message sent from each follower storage system (1138, 1150) to the leader storage system (1140). The reader will understand that, although many of the steps described above are shown and described as occurring in a particular order, no particular order is actually required. In practice, because the follower storage systems (1138, 1150) and the leader storage system (1140) are independent storage systems, each storage system may be performing some of the above-described steps in parallel. For example, one or more of the follower storage systems (1138, 1150) may receive (1116, 1142) information (1110) describing modifications to the dataset (1142), process (1118, 1144) the request (1104) to modify the dataset (1142), or acknowledge (1120, 1146) receipt of completion of the request (1104) to modify the dataset (1142) before processing the request (1104) to modify the dataset (1142). Alternatively, the leader storage system (1140) may have processed (1124) the request (1104) to modify the dataset (1142) before one or more of the follower storage systems (1138, 1150) have received (1116, 1142) the information (1110) describing the modifications to the dataset (1142), processed (1118, 1144) the request (1104) to modify the dataset (1142), or acknowledged (1120, 1146) receipt of the completion of the request (1104) to modify the dataset (1142).

[0307] The example method shown in FIG. 12 also includes determining (1158) by the leader storage system (1140) whether the request (1104) to modify the dataset (1142) has been processed (1118, 1144) by each of the follower storage systems (1138, 1150) before confirming (1134) receipt of completion of the request (1104) to modify the dataset (1142). The leader storage system (1140) may determine (1158) whether the request (1104) to modify the dataset (1142) has been processed (1118, 1144) by each of the follower storage systems (1138, 1150), for example, by determining whether the leader storage system (1140) has received an acknowledgment message or other message from each of the follower storage systems (1138, 1150) indicating that the request (1104) to modify the dataset (1142) has been processed (1118, 1144) by each of the follower storage systems (1138, 1150). In such an example, if the leader storage system (1140) determines in the affirmative (1162) that the request (1104) to modify the dataset (1142) has been processed (1118, 1144) by each of the follower storage systems (1138, 1150) and has also been processed (1124) by the leader storage system (1138), the leader storage system (1140) may proceed by acknowledging (1134) receipt of completion of the request (1104) to modify the dataset (1142) from the host (1102) that initiated the request (1104) to modify the dataset (1142). However, if the leader storage system (1140) determines that the request (1104) to modify the data set (1142) has not been processed (1160) (1118, 1144) by at least one of the follower storage systems (1138, 1150) or has not been processed (1124) by the leader storage system (1138), the leader storage system (1140) determines that the request (1104) to modify the data set (1142) has not been processed (1118, 1144) by at least one of the follower storage systems (1138, 1150) or has not been processed (1124) by the leader storage system (1138), the leader storage system (1140) determines that the request (1104) to modify the data set (1142) has been processed (1124) by all storage systems across which the data set (1142) is synchronously replicated. Since receipt of completion of the request (1104) to modify the dataset (1142) may be acknowledged (1134) to the host (1102) that initiated the request (1104) to modify the dataset (1142) only once it has been successfully processed by the storage systems (1138, 1140, 1150), the leader storage system (1140) may not yet acknowledge (1134) receipt of completion of the request (1104) to modify the dataset (1142) to the host (1102) that initiated the request (1104) to modify the dataset (1142).

[0308] The reader will understand that while the example method shown in FIG. 12 illustrates an embodiment in which a data set (1142) is synchronously replicated across three storage systems, one of which is a leader storage system (1140) and the remaining storage systems are follower storage systems (1138, 1150), other embodiments may include additional storage systems. In such other embodiments, the additional follower storage systems may operate in the same manner as the follower storage systems (1138, 1150) shown in FIG. 12.

[0309] For further explanation, FIG. 13 sets forth a flowchart illustrating an example method for servicing I/O operations directed to a dataset (1142) synchronized across multiple storage systems (1138, 1140) in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (1138, 1140) illustrated in FIG. 13 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 13 may include the same components as the storage systems described above, fewer components, or additional components.

13 includes receiving (1302), by a follower storage system (1138), a request (1104) to modify a dataset (1142). The request (1104) to modify a dataset (1142) may be implemented, for example, as a request to write the data to a location in the storage system (1138) that contains the data included in the dataset (1142), as a request to write the data to a volume that contains the data included in the dataset (1142), or as some other operation that causes a change to some portion of the data included in the dataset (1142). In the example method shown in FIG. 13, the request (1104) to modify a dataset (1142) is issued by a host (1102), which may be implemented, for example, as an application running on a virtual machine, as an application running on a computing device connected to the storage system (1138), or as some other entity configured to access the storage system (1138).

13 also includes transmitting (1304) a logical description (1306) of the request (1104) to modify the dataset (1142) from the follower storage system (1138) to the leader storage system (1140). In the example method shown in FIG. 13, the logical description (1306) of the request (1104) to modify the dataset (1142) may be formatted in a manner understood by the leader storage system (1138) and may include information describing the type of operation (e.g., a read-type operation, a snapshot-type operation) requested in the request (1104) to modify the dataset (1142), information describing where the I/O payload is located, information describing the size of the I/O payload, or some other information. In an alternative embodiment, the follower storage system (1138) may simply forward some portion (including all) of the request (1104) to modify the dataset (1142) to the leader storage system (1140).

13 also includes generating (1308), by the leader storage system (1140), information (1310) describing modifications to the dataset (1142). The leader storage system (1140) may generate (1308) the information (1310) describing modifications to the dataset (1142), for example, by determining ordering versus any other operations that are in progress, computing distributed state changes, for example, to common elements of metadata, across all members of a pod (e.g., all storage systems across which the dataset is synchronously replicated), etc. The information (1310) describing modifications to the dataset (1142) may be implemented, for example, as system-level information used to describe I/O operations performed by the storage system. The leader storage system (1140) may generate (1308) information (1310) describing modifications to the dataset (1142) by processing the request (1104) to modify the dataset (1142) sufficiently to understand what must occur in order to service the request (1104) to modify the dataset (1142). For example, the leader storage system (1140) may determine whether some ordering of execution of the request (1104) to modify the dataset (1142) relative to other requests to modify the dataset (1142) is required to produce equivalent results at each storage system (1138, 1140).

[0313] Consider an example in which request (1104) to modify dataset (1142) is implemented as a request to copy blocks from a first address range within dataset (1142) to a second address range within dataset (1142). In such an example, assume that three other write operations (Write A, Write B, Write C) are directed to the first address range of dataset (1142). In such an example, if the leader storage system (1140) services Write A and Write B (but does not service Write C) before copying blocks from a first address range in the data set (1142) to a second address range in the data set (1142), then the follower storage system (1138) must also instruct Write A and Write B (but not Write C) before copying blocks from the first address range in the data set (1142) to the second address range in the data set (1142) in order to produce consistent results. Thus, when the leader storage system (1140) generates (1308) information (1310) describing modifications to the data set (1142), in this example, the leader storage system (1140) could generate information (e.g., sequence numbers for Write A and Write B) that identifies other operations that must be completed before the follower storage system (1138) can process the request (1104) to modify the data set (1142).

[0314] The reader will further understand that the correct outcome for any operation must be committed to a point where it is recoverable before receipt of the operation can be acknowledged. However, in some cases, multiple operations can be committed together, or operations can be partially committed if recovery ensures correctness. For example, a snapshot could be committed locally with dependencies recorded on expected writes A and B, but A or B itself may not be committed. If receipt of the snapshot cannot be acknowledged and lost I/O cannot be recovered from another array, recovery may end up canceling the snapshot. Also, if write B overlaps with write A, the leader storage system may "instruct" B to be behind A, but A is actually discarded, and then the operation writing A will simply wait for B. Writes A, B, C, and D combined with snapshots between A, B and C, D will be able to commit and/or acknowledge receipt of some or all portions together, as long as recovery does not cause snapshot inconsistencies across the array, and as long as acknowledgments do not complete a later operation before an earlier operation has persisted to the point where it is guaranteed to be recoverable.

13 also includes transmitting (1312) information (1310) describing the modifications to the dataset (1142) from the leader storage system (1140) to the follower storage system (1338). The transmitting (1312) information (1310) describing the modifications to the dataset (1142) from the leader storage system (1140) to the follower storage system (1338) may be performed, for example, by the leader storage system (1140) sending one or more messages to the follower storage system (1138). However, given the fact that the follower storage system (1138) was the original recipient of the request (1104) to modify the dataset (1142), the leader storage system (1140) may not need to transmit the I/O payload of the request (1104) to modify the dataset (1142). Thus, the follower storage system (1138) may extract the I/O payload from the request (1104) to modify the dataset (1142), the follower storage system (1138) may receive the I/O payload as part of one or more other messages associated with the request (1104) to modify the dataset (1142), and the follower storage system (1138) may be able to access the I/O payload because it is stored in a known location (e.g., a buffer in the follower storage system (1138) accessed via RDMA or RDMA-like access) or in some other manner by the host.

13 also includes processing (1318) the request (1104) to modify the dataset (1142) by the leader storage system (1140). In the example method shown in FIG. 13, the leader storage system (1140) may process (1318) the request (1104) to modify the dataset (1142), for example, by modifying the contents of one or more storage devices (e.g., NVRAM devices, SSDs, HDDs) included in the leader storage system (1140) according to information (1110) describing the modifications to the dataset (1142) as well as the I/O payload received from the follower storage system (1138). Consider an example in which the request (1104) to modify the dataset (1142) is implemented as a write operation directed to a volume included in the dataset (1142), and the information (1110) describing the modification to the dataset (1142) indicates that the write operation can be performed only after a previously issued write operation has been processed. In such an example, processing (1318) the request (1104) to modify the dataset (1142) may be implemented by the leader storage system (1140) first verifying that the previously issued write operation has been processed at the leader storage system (1140), and then writing the I/O payload associated with the write operation to one or more storage devices included in the leader storage system (1140). In such an example, the request (1104) to modify the dataset (1142) may be considered completed and successfully processed, for example, when the I/O payload has been committed to persistent storage in the leader storage system (1140).

[0317] The example shown in FIG. 13 also includes acknowledging (1320) by the leader storage system (1140) to the follower storage system (1138) receipt of completion of the request (1104) to modify the dataset (1142). In the example shown in FIG. 13, the leader storage system (1140) may acknowledge (1320) receipt of completion of the request (1104) to modify the dataset (1142), for example, by using one or more acknowledgment (1322) messages sent from the leader storage system (1140) to the follower storage system (1138), or via some other suitable mechanism.

13 also includes receiving (1314) information (1310) from the leader storage system (1140) describing modifications to the data set (1142). The follower storage system (1138) may receive (1314) the information (1110) describing modifications to the data set (1142) from the leader storage system (1140), for example, via one or more messages sent from the leader storage system (1140) to the follower storage system (1138). The one or more messages may be sent from the leader storage system (1140) to the follower storage system (1138) over one or more dedicated data communication links between the two storage systems (1138, 1140) by the leader storage system (1140) writing the messages to predetermined storage locations (e.g., queue locations) on the follower storage system (1138) using RDMA or similar mechanisms or otherwise. However, the reader will understand that in the example method shown in FIG. 13, the follower storage system (1138) can extract the relevant I/O payload from the request (1104) to modify the dataset (1142) received by the follower storage system (1138), the follower storage system (1138) can extract the relevant I/O payload from one or more other messages received from the host (1102), and the leader storage system (1140) need not send the I/O payload associated with the request (1104) to modify the dataset (1142) to the follower storage system (1138) because the follower storage system (1138) can obtain the I/O payload in some other manner due to the fact that the follower storage system (1138) was the target of the request (1104) to modify the dataset (1142) issued by the host (1102).

[0319] In one embodiment, the follower storage system (1138) may receive information (1110) describing modifications to the dataset (1142) from the leader storage system (1140) using SCSI requests (write from sender to receiver, or read from receiver to sender) as the communication mechanism (1314). In such an embodiment, a SCSI write request is used to encode the information intended for transmission (including any data and metadata), which may be delivered to a special pseudo-device, or over a specially configured SCSI network, or through any other agreed-upon addressing mechanism. Alternatively, the model may use a special device, a specially configured SCSI network, or other agreed-upon mechanism to issue a set of open SCSI read requests from the receiver to the sender. The encoded information, including the data and metadata, is delivered to the receiver in response to one or more of these open SCSI requests. Such a model may be implemented over a Fibre Channel SCSI network, which is often deployed as a "dark fibre" storage network between data centers. Such a model also allows for multipathing from the host to the remote array and the use of the same network lines for bulk inter-array communication.

13 also includes processing (1316) the request (1104) to modify the dataset (1142) by the follower storage system (1138). In the example method shown in FIG. 13, the follower storage system (1138) may process (1316) the request (1104) to modify the dataset (1142) by modifying the contents of one or more storage devices (e.g., NVRAM devices, SSDs, HDDs) included in the follower storage system (1138) in accordance with the information (1110) describing the modifications to the dataset (1142). Consider an example in which the request (1104) to modify the dataset (1142) is implemented as a write operation directed to a volume included in the dataset (1142), and the information (1110) describing the modifications to the dataset (1142) indicates that the write operation can be performed only after a previously issued write operation has been processed. In such examples, processing (1316) the request (1104) to modify the dataset (1142) may be performed by the follower storage system (1138) first verifying that a previously issued write operation has been processed at the follower storage system (1138), and then writing the I/O payload associated with the write operation to one or more storage devices included in the follower storage system (1138). In such examples, the request (1104) to modify the dataset (1142) may be considered completed and successfully processed, for example, when the I/O payload associated with the request (1104) to modify the dataset (1142) has been committed to persistent storage in the follower storage system (1138).

13 also includes receiving (1324) an indication from the leader storage system (1140) that the leader storage system (1140) has processed (1104) the request to modify the data set (1142). In this example, the indication that the leader storage system (1140) has processed (1104) the request to modify the data set (1142) is embodied as an acknowledgment (1322) message sent from the leader storage system (1140) to the follower storage system (1138). The reader will understand that although many of the steps described above are shown and described as occurring in a particular order, no particular order is actually required. In practice, the follower storage system (1138) and the leader storage system (1140) are independent storage systems, and therefore each storage system may be performing some of the steps described above in parallel. For example, the follower storage system (1138) may receive (1324) an indication from the leader storage system (1140) that the leader storage system (1140) processed (1316) the request (1104) to modify the data set (1142). Similarly, the follower storage system (1138) may receive (1324) an indication from the leader storage system (1140) that the leader storage system (1140) processed (1314) the request (1104) to modify the data set (1142) before the leader storage system (1140) received (1314) information (1110) from the leader storage system (1140) describing the modifications to the data set (1142).

13 also includes acknowledging (1326) receipt of completion of the request (1104) to modify the dataset (1142) by the follower storage system (1138). Acknowledging (1326) receipt of completion of the request (1104) to modify the dataset (1142) may be performed, for example, by the follower storage system (1138) issuing an acknowledgment (1328) message to the host (1102) that issued the request (1104) to modify the dataset (1142). In the example method shown in FIG. 13, the follower storage system (1138) may determine whether the request (1104) to modify the dataset (1142) has been processed (1318) by the leader storage system (1140) before acknowledging (1328) receipt of completion of the request (1104) to modify the dataset (1142). The follower storage system (1138) may determine whether the request (1104) to modify the dataset (1142) has been processed (1318) by the leader storage system (1140), for example, by determining whether the follower storage system (1138) has received an acknowledgment message or other message from the leader storage system (1140) indicating that the request (1104) to modify the dataset (1142) has been processed (1318) by the leader storage system (1140). In such an example, if the follower storage system (1138) affirmatively determines that the request (1104) to modify the dataset (1142) has been processed (1318) by the leader storage system (1140) and the follower storage system (1138) has also processed (1316) the request (1104) to modify the dataset (1142), the follower storage system (1318) may proceed by acknowledging (1326) receipt of completion of the request (1104) to modify the dataset (1142) from the host (1102) that initiated the request (1104) to modify the dataset (1142). However, if the leader storage system (1140) determines (1318) that the request (1104) to modify the dataset (1142) has not been processed by the leader storage system (1140) or that the follower storage system (1138) has not yet processed (1316) the request (1104) to modify the dataset (1142), the follower storage system (1138) determines that the request (1104) to modify the dataset (1142) has not yet been processed by the leader storage system (1140) in all of the data sets in which the dataset (1142) is synchronously replicated. Since receipt of completion of the request (1104) to modify the dataset (1142) may be acknowledged to the host (1102) that initiated the request (1104) to modify the dataset (1142) only when the request has been successfully processed by all storage systems (1138, 1140), the follower storage system (1138) may not yet acknowledge (1326) receipt of completion of the request (1104) to modify the dataset (1142) to the host (1102) that initiated the request (1104) to modify the dataset (1142).

[0323] For further explanation, FIG. 14 sets forth a flowchart illustrating an example method for servicing I/O operations directed to a dataset (1142) synchronized across multiple storage systems (1138, 1140, 1334) in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (1138, 1140, 1334) illustrated in FIG. 13 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 13 may include the same components as the storage systems described above, fewer components, or additional components.

14 may be similar to the example method shown in FIG. 13. The example method shown in FIG. 4 includes receiving (1302) a request (1104) to modify a dataset (1142) by a follower storage system (1138), transmitting (1304) a logical description (1306) of the request (1104) to modify the dataset (1142) from the follower storage system (1138) to a leader storage system (1140), generating (1308) information describing the modifications to the dataset (1142) by the leader storage system (1140), processing (1318) the request (1104) to modify the dataset (1142) by the leader storage system (1140), and transmitting (1318) the request (1104) to the follower storage system (1138) to the leader storage system (1140) to the follower storage system (1138). The method may be similar to the example method shown in FIG. 13 , as it also includes acknowledging receipt of completion of the request (1104) to modify the dataset (1142) (1320), receiving information (1310) from the leader storage system (1140) describing the modifications to the dataset (1142) (1314), processing (1316) the request (1104) to modify the dataset (1142) by the follower storage system (1138), receiving (1324) an indication from the leader storage system (1140) that the leader storage system (1140) has processed the request (1104) to modify the dataset (1142), and acknowledging receipt of completion of the request (1104) to modify the dataset (1142) (1326) by the follower storage system (1138).

[0325] However, the example method shown in FIG. 14 differs from the example method shown in FIG. 13 because it illustrates an embodiment in which a data set (1142) is synchronously replicated across three storage systems, one of which is a leader storage system (1140) and the remaining storage systems are follower storage systems (1138, 1134). In such an example, the additional follower storage system (1134) performs many of the same steps as the follower storage system (1138) shown in FIG. 13, since it can receive (1142) information (1110) from the leader storage system (1140) describing modifications to the data set (1142) and process (1142) requests (1104) to modify the data set (1142) in accordance with the information (1110) describing modifications to the data set (1142).

[0326] In the example method shown in FIG. 14, the leader storage system (1140) can send (1338) information (1110) describing modifications to the dataset (1142) to all of the follower storage systems (1138, 1134). In the example method shown in FIG. 14, the additional follower storage system (1334) can also acknowledge (1330) receipt of completion of the request (1104) to modify the dataset (1142) to the follower storage system (1138) that received (1302) the request (1104) to modify the dataset (1142). In the example method shown in FIG. 14, the additional follower storage system (1334) may acknowledge (1330) receipt of completion of the request (1104) to modify the dataset (1142) to the follower storage system (1138) that received (1302) the request (1104) to modify the dataset (1142), e.g., by use of one or more acknowledgment (1332) messages sent from the additional follower storage system (1334) to the follower storage system (1138) that received (1302) the request (1104) to modify the dataset (1142), or via some other suitable mechanism.

14, the follower storage system (1138) that received (1302) the request (1104) to modify the dataset (1142) may also receive (1336) an indication that all other follower storage systems (1134) have processed the request (1104) to modify the dataset (1142). In this example, the indication that all other follower storage systems (1134) have processed the request (1104) to modify the dataset (1142) is embodied as an acknowledgment (1332) message sent from the other follower storage systems (1134) to the follower storage system (1138) that received (1302) the request (1104) to modify the dataset (1142). The reader will understand that although many of the steps described above are shown and described as occurring in a particular order, no particular order is actually required. In practice, because the follower storage systems (1138, 1334) and the leader storage system (1140) are independent storage systems, each storage system may be performing some of the above-described steps in parallel. For example, the follower storage system (1138) may receive (1324) an indication from the leader storage system (1140) that the leader storage system (1140) has processed (1316) the request (1104) to modify the data set (1142). Furthermore, the follower storage system (1138) may receive (1336) an indication that the request (1104) to modify the data set (1142) has been processed (1324) by all other follower storage systems (1334) before receiving (1324) an indication that the leader storage system (1140) has processed (1104) the request to modify the data set (1142). Alternatively, the follower storage system (1138) may receive (1336) an indication that it has processed (1316) the request (1104) to modify the data set (1142) before all other storage systems (1334) have processed (1316) the request (1104) to modify the data set (1142). Similarly, the follower storage system (1138) may receive (1324) an indication from the leader storage system (1140) that it has processed (1324) the request (1104) to modify the data set (1142) before the leader storage system (1140) receives (1314) information (110) from the leader storage system (1140) describing the modifications to the data set (1142). Additionally, the follower storage system (1138) may receive (1336) an indication that all other follower storage systems (1334) have processed (1104) the request to modify the data set (1142) before receiving (1314) the information (1110) describing the modifications to the data set (1142) from the leader storage system (1140).

[0328] Although not explicitly shown in FIG. 14, the follower storage system (1138) may determine whether the request (1104) to modify the dataset (1142) has been processed (1318) by the leader storage system (1140) and also processed (1144) by all other follower storage systems (1334) before acknowledging receipt of completion of the request (1104) to modify the dataset (1142) (1328). The follower storage system (1138) may determine whether the request (1104) to modify the dataset (1142) has been processed (1318) by the leader storage system (1140) and also processed (1144) by all other follower storage systems (1334), for example, by determining whether the follower storage system (1138) has received acknowledgment messages from the leader storage system (1140) and all other follower storage systems (1334) indicating that the request (1104) to modify the dataset (1142) has been processed (1318, 1144) by each storage system (1140, 1334). In such an example, if the follower storage system (1138) affirmatively determines that the request (1104) to modify the dataset (1142) has been processed by the leader storage system (1140), all other follower storage systems (1334), and the follower storage system (1138), the follower storage system (1138) may proceed by acknowledging (1326) receipt of completion of the request (1104) to modify the dataset (1142) to the host (1102) that initiated the request (1104) to modify the dataset (1142). However, if the leader storage system (1140) determines that the request (1104) to modify the data set (1142) has not been processed by the leader storage system (1140), all other follower storage systems (1334), or at least one of the follower storage system (1138), the follower storage system (1138) determines that the request (1104) to modify the data set (1142) has not been processed by all storage systems (1334) across which the data set (1142) is synchronously replicated. Since receipt of completion of the request (1104) to modify the dataset (1142) may be acknowledged (1134) to the host (1102) that initiated the request (1104) to modify the dataset (1142) only when the request has been successfully processed by the host (1102) that initiated the request (1104) to modify the dataset (1142), the follower storage system (1138) may not yet acknowledge (1326) receipt of completion of the request (1104) to modify the dataset (1142) to the host (1102) that initiated the request (1104) to modify the dataset (1142).

[0329] Although not explicitly shown in FIG. 14, in some embodiments, a follower storage system (1138) that receives (1302) a request (1104) to modify a dataset (1122) in an effort to unlock any concurrent overlapping reads in progress at one or more storage systems (1138, 1140, 1334) may send a message back to the leader storage system (1140) and to other follower storage systems (1334) to signal that the modification operation is complete wherever possible. Alternatively, a follower storage system (1138) that receives a request (1104) to modify a dataset (1122) may send that message to the leader storage system (1138), which could then send a message to propagate the completion and remove any obstructions to the reads wherever possible.

[0330] The reader will understand that although the example method illustrated in FIG. 14 depicts an embodiment in which a data set (1142) is synchronously replicated across three storage systems, one of which is a leader storage system (1140) and the remaining storage systems are follower storage systems (1138, 1334), other embodiments may include additional storage systems. In such other embodiments, the additional follower storage systems may operate similarly to the other follower storage system (1334) illustrated in FIG. 14.

[0331] The reader will also understand that while only the example shown in FIG. 12 explicitly illustrates an embodiment in which information (1310) describing modifications to dataset (1142) includes ordering information (1152) for the request (1104) to modify dataset (1142), common metadata information (1154) associated with the request (1104) to modify dataset (1142), and I/O payload (1114) associated with the request (1104) to modify dataset (1142), information describing modifications to dataset (1142) (1310) may include all (or a subset) of the information associated with the examples shown in the remaining figures. Additionally, in embodiments in which the request (1104) to modify the dataset (1142) includes a request to take a snapshot of the dataset (1142), the information (1310) describing the modifications to the dataset (1142) may also include identification of one or more other requests to modify the dataset (1142) that are included in the contents of the snapshot of the dataset (1142) in each of the figures described above.

[0332] The reader will understand that several situations can be accommodated by having information (1310) describing modifications to dataset (1142) include identification of one or more other requests modifying dataset (1142) that are included in the contents of a snapshot of dataset (1142), rather than including information identifying one or more other requests modifying dataset (1142) that must be completed before the snapshot is taken. One situation is where an atomic operation performs a snapshot and the last few writes can be completed in the same atomic update, i.e., the last few writes do not complete "before" the snapshot. Another situation is where writes may actually complete after the snapshot point is taken, if the writes are included when they complete, and if the snapshot itself is not considered complete until all writes have been completed by all syncing storage systems. Finally, writes that were not indicated to the requester as completed before the snapshot was received may be included or omitted from the snapshot as a result of recovery actions. Essentially, recovery can rewrite the detailed history of received operations, provided the outcome is consistent and does not violate any guarantees associated with the operations signaled as completed.

[0333] For additional explanation, FIG. 15 sets forth a flowchart illustrating an example method for mediating between storage systems that synchronously replicate a dataset in accordance with some embodiments of the present disclosure. While the example method illustrated in FIG. 15 illustrates an embodiment in which dataset (1512) is synchronously replicated across only two storage systems (1514, 1524), the example illustrated in FIG. 15 can be extended to embodiments in which dataset (1512) is synchronously replicated across additional storage systems.

[0334] In the following example, mediation between a collection of storage systems (1514, 1524) for a pod allows a storage system to resolve lost communication with a paired system, where communication may be lost due to a communication failure or some other type of system failure. As described below, the mediation solution may include the use of a quorum and an external control system to determine which storage system should continue processing I/O operations directed to the pod dataset and competing for, for example, a mediator. However, the advantage of mediation is that it is simpler than a quorum protocol, and mediation works well with a common configuration, two storage system configurations for synchronously replicated storage systems. Furthermore, mediation may be more robust and easier to configure than external control systems and many other types of resources that may be competed for.

[0335] As shown in FIG. 15, multiple storage systems (1514, 1524) synchronously replicating a dataset (1512) may communicate with an intermediary service (1500) over a network (1554)—the intermediary service (1500) may determine which storage system continues to service the dataset in the event of a communication failure between the storage systems, if a storage system goes offline, or due to some other triggering event. Intermediation is advantageous because if the storage systems cannot communicate with each other, they may not be able to maintain the synchronously replicated dataset, and any received requests to modify the dataset would otherwise be unavailable for service because the dataset would become unsynchronized. In this example, intermediary services for the storage systems synchronously replicating the dataset may be provided by an intermediary service (1500) that is external to the storage systems (1514, 1524). While only two storage systems (1514, 1524) are shown in this example, generally, some other number of two or more storage systems may be part of the in-sync list that are synchronously replicating the dataset. Specifically, if a first storage system (1514) detects a triggering event, such as the loss of a communication link (1516) to a second storage system (1524), the first storage system (1514) may contact the external intermediary service (1500) to determine whether it can successfully take over the task of removing the non-communicating storage system from the in-sync list that specifies the storage systems with which it is in sync for replicating the dataset. In other cases, the first storage system (1514) may contact the external intermediary service (1500) to determine that it, i.e., the first storage system (1500), may have been removed from the in-sync list by the second storage system. In these examples, under normal conditions, the storage systems (1514, 1524) do not need to be in continuous communication with the external intermediary service (1500) because they do not require any information from the intermediary service (1500) to operate properly and to maintain a synchronized replication of the dataset (1512). In other words, in these examples, the intermediary service (1500) does not have an active role in managing the membership of the in-sync list; further, the intermediary service (1500) may not even be aware of the normal operation of the storage systems (1514, 1524) on the in-sync list. Instead, the intermediary service (1500) may simply provide persistent information used by the storage systems (1514, 1524) to determine membership of the in-sync list or to determine whether a storage system can act to detach another storage system.

[0336] In some examples, the intermediary service (1500) may be contacted by one or more storage systems (1514, 1524) in response to a triggering event, such as a communication link failure that prevents the storage systems (1514, 1524) from communicating with each other; however, each storage system (1514, 1524) may be able to communicate with the intermediary service (1500) via a communication channel that is different from the communication channel used between the storage systems (1514, 1524). Consequently, the storage systems (1514, 1524) may not be able to communicate with each other, but the storage systems (1514, 1524) may still communicate with the intermediary service (1500), and the storage systems (1514, 1524) may use the intermediary service (1500) to determine which storage system may proceed to service the data storage request. Additionally, a storage system that wins mediation from the intermediary service (1500) may detach another storage system and update its in-sync list to indicate storage systems that may continue to synchronously replicate the dataset (1512). In some examples, the intermediary service (1500) may process various types of requests, such as a request to set a membership list that includes the requester storage system and excludes another storage system. In this example, if the intermediary service (1500) currently lists the requester as a member, the request completes successfully; if the intermediary service (1500) does not currently list the requester as a member, the request fails. In this manner, if two storage systems (1514, 1524) are each making requests at approximately the same time, and the requests serve to exclude the other, then the first request received may succeed—the intermediary service sets the membership list to exclude the other storage system according to the first request—and the second request received may fail because the membership list is set to exclude it. Mutually exclusive access to the shared resource that stores the membership list helps ensure that only one system at a time is allowed to configure the membership list.

[0337] In another example, mediation may be based on a specific identifier whose value indicates a pod membership partition identifier and can be defined to assert membership partitioning or removal of any collection of storage systems from a pod. A 'pod,' as the term is used here and throughout the rest of this application, may be implemented as a management entity representing a dataset, a collection of managed objects and management operations, a collection of access operations to modify or read the dataset, and multiple storage systems. Such management operations may modify or query managed objects equally well through any of the storage systems, with access operations to read or modify the dataset operating equally well through any of the storage systems. Each storage system may store a separate copy of the dataset as an appropriate subset of the dataset stored and advertised for use by the storage system, and a managed object or operation to modify the dataset performed and completed through any one storage system is reflected in a subsequent managed object to query the pod or in a subsequent access operation to read the dataset. Additional details regarding the 'pod' may be found in previously filed provisional patent application Ser. No. 62/518,071, which is incorporated herein by reference.

[0338] The partition identifier may be local information stored on a given storage system in addition to the storage system storing the pod membership list. Systems that are properly communicating and synchronized with each other may have the same partition identifier, and when a storage system is added to a pod, the current partition identifier may then be copied along with the pod data contents. In this example, when one set of storage systems is not communicating with another set of storage systems, one storage system from each set may attempt to set the partition identifier on a shared resource maintained by the intermediary service (1500) by coming up with a new, unique partition identifier and using a specific operation that succeeds for the storage system that first obtains a lock on the shared resource; if the other storage system fails to obtain a lock on the shared resource, the attempt to perform the specific operation fails. In one embodiment, an atomic compare-and-set operation may be used, and the last partition identifier value stored by the intermediary service (1500) may be provided by the storage system to have permission to change the partition identifier to a new value. In this example, the compare-and-set operation may succeed for storage systems that are aware of the current partition identifier—the storage system that initially sets the partition identifier value will be the storage system that is aware of the current partition identifier value. Additionally, a conditional store or PUT operation, which may be available in a Web Services protocol, may function to set the partition identifier, as described in this example. In other cases, such as in a SCSI environment, a compare-and-write operation may be used. In still other cases, the intermediary service (1500) may perform the compare-and-set operation by receiving a request from a storage system, the request also indicating the old and new partition identifier values, and if and only if the currently stored value is equal to the old partition identifier, the intermediary service (1500) changes the stored partition identifier to the new partition identifier.

[0339] In this manner, intermediation based on the partition identifier may be used to persist information that may be used by a storage system to determine whether a given storage system is included in a bounded set of consistent pod members. In some cases, the partition identifier may change only in the event of a spontaneous detachment due to a failure of either the storage system or the network interconnect. In these examples, a storage system that takes itself offline for a pod in a controlled manner may communicate with other storage systems to remove itself as an in-sync pod member, thus not requiring the creation of a new mediated pattern identifier. Furthermore, a storage system that removes itself as a member of an in-sync pod may then add itself back as an in-sync pod member in a controlled manner that does not require a new mediated partition identifier. Furthermore, new storage systems may be added to an in-sync pod if the storage system is in communication with the in-sync pod member, and the new storage systems may add themselves in a controlled manner that does not require a new mediated partition identifier.

[0340] Consequently, an advantage of the mediated partition identifier mechanism is that the intermediary service (1500) may be needed only when there is a failure or other triggering event in which at least one set of storage systems reacts by attempting to remove one or more non-communicating storage systems from the in-sync pod membership list; non-communicating storage systems may attempt to do the same, but vice versa. Another advantage is that the intermediary service (1500) can never be absolutely reliable and can have little impact on the availability of the overall storage service provided by the in-sync pod members. For example, if two synchronously replicated storage systems each fail once a year, then the first of the two storage systems should fail and the second storage system should successfully mediate to remove the first storage system, unless the intermediary service (1500) is unavailable at that precise moment. In short, if the intermediary service (1500) is operational and available at least 99% of the time, the probability that the intermediary service (1500) will be unavailable when needed is extremely low. In this example, there would only be a 1 in 100 (less than 1%) chance that the intermediary service (1500) would be unavailable at a critical time - this could reduce a once-a-year outage to a once-a-century outage. However, to reduce the likelihood of the intermediary service (1500) being unavailable, the intermediary service (1500) may be periodically monitored to alert an administrator if the intermediary service is generally unavailable, and the intermediary service (1500) may monitor storage systems to generate alerts if a particular storage system becomes unavailable.

[0341] In another example, as an alternative to using the partition identifier associated with a synchronizing member for a pod, the mediation service (1500) may provide a one-time mediation race target. Specifically, a mediation race target may be established each time a synchronizing member storage system for a pod needs to allow for the possibility that one storage system may be detached by another. For example, the agreed-upon key in the mediation value table may be set to a new value one time; to win mediation, a storage system sets the agreed-upon key to a unique value that other competing storage systems will not use. Prior to the mediation race, an agreed-upon key may not exist; if an agreed-upon key exists, the key may be set to some agreed-upon precursor value, such as UNSET or a null value. In this example, an operation to set a key to a particular value succeeds if the key does not exist, if the key is in the UNSET state, or if the key is set to a value equal to the current value—otherwise, the operation to set the key fails. Once a set of storage systems wins the arbitration, the remaining set of storage systems may define a new key for future arbitration. In this example, before a storage system knows that it may have won the arbitration race, it may record the value it uses before the arbitration race so that it can reuse the value if it fails and recovers or reboots. If two or more storage systems are in communication and competing with each other for some other set of non-communicating storage systems, this value may be shared with those other communicating storage systems so that any one of them can continue the arbitration race and, perhaps after some additional series of failures, participate in a second arbitration race. For example, it may be necessary to compete for or verify the accuracy of the first arbitration race target before competing for a unique value for the second arbitration race target. Notably, this sequence may be necessary until the second arbitration race target has been reliably distributed to all storage systems sharing the first arbitration race target and all storage systems are made aware that it has been reliably distributed. At that point, there may be no ongoing need to first compete for the first intermediary target before competing for the second intermediary target.

[0342] In some examples, the intermediary service (1500) may be hosted on a computer system provided by an organization other than the organization or owner of the storage system being brokered. For example, if a vendor sells two storage systems to a customer, the vendor may host the mediator on a server provided in a vendor-owned or vendor-managed data center, or the vendor may contract with a cloud service provider to host the service. The vendor may also ensure that the intermediary service is sufficiently reliable and distinct from the customer's failure zone. In some cases, without excluding other cloud service providers, the intermediary service may be hosted on Amazon Web Services™, and the intermediary service may be implemented with DynamoDB for reliable database services, which may provide support for conditional store primitives as web API database updates. In some cases, the intermediary service may be implemented to operate across multiple cloud service provider domains or failure domains to further increase reliability. One advantage of using a vendor to provide the intermediary service is that the intermediary service is easy to configure. Furthermore, during pod creation, the storage system may obtain a cryptographic token from the intermediary service and store the cryptographic token in addition to storing the partition identifier and pod membership list—the cryptographic token can be used to securely communicate unique intermediary service information for the pod.

[0343] In some cases, the intermediary service (1500) may be unavailable when a storage system attempts intermediation, and the following method provides a process for recovering, at least eventually, from such a service outage. For example, if a first set of storage systems attempts to detach a second set of storage systems through the intermediary service, but the first set of storage systems cannot communicate with the intermediary service (1500), then the first set of storage systems cannot complete the detach operation and cannot continue to provide service to the pod. In some cases, if the two sets of storage systems manage to reconnect with each other so that all synchronizing storage systems are communicating again—although the intermediary service (1500) is still unavailable—the two sets of storage systems may synchronize the pod and resume providing service to the pod. However, in this example, one or more requests may have been sent to the intermediation service (1500) to change the partition identifier or any other property associated with the intermediation, and none of the storage systems may be sure whether the request was received or not received and processed, and acknowledgments may have been lost. As a result, if there is a collection of failed storage systems or network interconnects, then the storage systems may not be sure what value to assert for the partition identifier if and when the intermediation service (1500) comes back online. In such a situation, it is preferable for the pod's services to resume either when all synchronizing storage systems come back online and resume communication, or when a synchronizing storage system is able to reconnect to the intermediation service (1500). In one embodiment, when all synchronizing storage systems reconnect, they all exchange known partition identifier values that may have been sent to the intermediation service (1500). For example, if two storage systems each attempt to change a partition identifier value, where one storage system attempts to change the partition identifier to, for example, 1749137481890 and another storage system attempts to change the partition identifier to, for example, 87927401839, and the last value known to have been acknowledged by the intermediary service (1500) was 79223402936, then the intermediary service (1500) may currently store any of these three partition identifier values. As a result, future attempts to change the intermediary partition identifier to a new value may supply any or all of these three partition identifiers in an attempt to gain authorization to make the change. Furthermore, a fourth attempt to change the partition identifier value may also encounter a failure, resulting in a fourth value that may need to be remembered by any storage system that attempts yet another intermediation at a later time. Furthermore, if any storage system successfully changes the intermediary service (1500) partition identifier value, that storage system may purge the older partition identifier value from any currently synchronizing storage systems and from any storage systems that will be synchronizing in the future.

[0344] In another example, the mediation service (1500) may mediate based on a unique key prepared for each potential future competition. In such a case, the synchronizing storage systems may agree to use the new key. Given that the new key may not be set atomically on all storage systems simultaneously until all synchronizing storage systems have received and recorded the new key, all storage systems should retain the old key and value that each storage system attempted to set in any previous mediation attempt. In this example, any earlier uncontested keys and any earlier key/value mediation attempts may be circulated among all synchronizing storage systems for the pod and recorded at each such storage system, along with the new key to use in future mediation attempts. For each earlier uncontested key (not including the new key), this exchange may select a single agreed-upon value that all systems may use in a competition for that key. After all synchronizing systems for the pod have received and recorded all of these mediation keys and values (and the new agreed-upon key for any future competitions), the storage systems in the pod may then agree to select a single new key and discard the older keys. Note that two or more storage systems may attempt to set the same brokered key to different values, and all such values may be recorded. If there is a failure during the process of exchanging or receiving all these brokered keys and key/value pairs for past brokering attempts, then some storage systems may not receive and record the new brokered key and value, while others may have received and recorded them. If the brokering service (1500) becomes available before all in-sync storage systems for a pod are able to reconnect with each other, then a subset of the storage systems for the pod may attempt to use the brokering service (1500) to detach another storage system from the pod. To win the brokerage, a storage system may attempt to set all recorded keys to their recorded values, and if that is successful, then attempt to set a new key to a unique value. If two or more values were recorded for the same key, then the step succeeds if setting any one of those values is successful. If the first step (setting the previous key) fails, or the second step (setting the new key to a new, unique value) fails, then the storage systems participating in that attempt to broker may go offline (retaining the value it attempted to set for the new key). If both steps succeed, then the communicating storage system may detach the non-communicating storage system and continue serving the pod. As an alternative to exchanging all past keys and values, a storage system may record only the keys and values it tries, without exchanging keys and values from other storage systems for the pod. Then, when a synchronizing storage system reconnects with other synchronizing storage systems for the pod (none of which were successful in interacting with the brokering service), the synchronizing storage systems may exchange one new brokering key and then exchange acknowledgments that they both received and recorded the agreed-upon new key. If a failure prevents the exchange of acknowledgments, then future attempts to broker (to the currently available brokering service) by storage systems that did not receive the new key may attempt to re-assert its previous key and value. A storage system that receives a new key but does not receive an indication that all storage systems for the pod have received the key may assert its previous brokered key, rather than just asserting a value for the new key, first the previous key, then the new key. Its future brokering attempts may still fail, which may result in the storage system again reconnecting to other in-sync storage systems and again incompletely exchanging a new key, leading to another key. This adds another key. As the keys grow over time with a series of incomplete exchanges of new keys, future brokering attempts by the storage system may re-assert each of its keys in the order in which they were recorded, along with any values the storage system previously asserted for those keys, until the storage system has successfully asserted values for all keys, or the storage system experiences a key assertion failure, at which point the storage system stops asserting keys and goes offline.

[0345] In another example, a new mediation service may be configured when the current mediation service is unavailable. For example, if all in-sync storage systems for a pod are in communication with each other but not with the current mediation service, then the pod may be configured with a new mediation service. This is similar to the previous algorithm for selecting a new key or new mediation value, but the new key is further configured to use the new mediation service rather than simply being a different key associated with the same service. Furthermore, if there is a failure during this operation, as with the previous algorithm, some systems may compete for the older key, and therefore systems that know the old key and the new key with the new mediation service may compete for the new key with the new mediation service. If the previous mediation service is permanently unavailable, then all in-sync storage systems should eventually reconnect with each other and complete the exchange of the new mediation service, and any keys and values associated with the new mediation service, before the pod service can safely resume.

[0346] In another example, a model for resolving failures may be to implement priority rules to favor one storage system over another. In this example, if a preferred storage system is running, it remains running and detaches any storage systems with which it is not communicating. Additionally, any other systems that do not have proven communication with the preferred system take themselves offline. In this example, the preferred storage system detaches the reconnecting storage system, and the reconnecting storage system must then first be resynchronized to synchronize it for the pod before it can resume serving the pod, whereas when the non-preferred storage system eventually reconnects with the preferred storage system, if the preferred storage system had not yet detached the reconnecting storage system, the two storage systems may recover and resume from a state where both storage systems are in sync. While having a preferred storage system may not be useful for providing high availability, it may be useful for other uses of synchronous replication, particularly asymmetric synchronous replication. Take, for example, mirroring from a large central storage system within a data center or campus to a smaller (perhaps less managed) storage system running closer to the application servers, e.g., in a top-of-rack configuration. In this case, in the event of a network outage or top-of-rack storage system failure, it may be beneficial to always favor the larger, more centralized storage system while completely halting service for the pod if the centralized storage system fails. Such a top-of-rack storage system may be used solely to improve read performance or reduce load on the data center storage network, but if asynchronous replication or other data management services are running solely on the centralized system, it may be preferable to reroute traffic to the central storage system or take it out of service and request technical support, rather than allowing the top-of-rack storage system to continue on its own. Furthermore, the priority rules may be more complex—there may be two or more such “preferred” storage systems combined, perhaps with some number of additional storage systems relying on the preferred or required storage system. In this example, a pod is online if all preferred or required storage systems are running, and down if some are not. This is similar to a quorum model where the quorum size is the same as the number of voting members, but it is easier to implement than a generalized quorum model that allows for less than all voting members.

[0347] In another example, a combination of mechanisms may be used that may be useful when a pod is spread across three or more storage systems. In one example, priority rules may be combined with metadata. In a top-of-rack example, a larger, central storage system in a data center or campus may itself be synchronously replicated to a larger storage system in a second location. In that case, the top-of-rack storage system may never restart on its own and may give priority to either of the larger, central storage systems in the two locations. The two larger storage systems in that case may be configured to mediate between each other, such that any smaller storage system that can connect to either of the two larger storage systems that remain online may continue to serve its pod, and any smaller storage system that cannot connect to either of the two larger storage systems (or can only connect to one that is offline for the pod) may cease to serve the pod. Furthermore, the priority model may be combined with a quorum-based model. For example, three large storage systems in three locations might use a quorum model between each other, with smaller satellite or top-of-rack storage systems lacking any votes and only functioning if they can connect to one of the larger, synchronizing storage systems that is online.

[0348] In another example of combining mechanisms, brokering may be combined with a quorum model. For example, there may be three storage systems that vote among themselves to ensure that a third storage system, with which two storage systems normally do not communicate, can be safely detached. On the other hand, one storage system cannot absolutely detach two other storage systems on its own. However, after two storage systems successfully detach the third storage system, the configuration is now reduced to two storage systems that agree that they are in sync and agree on the fact that the third storage system is detached. In that case, the remaining two storage systems may agree to handle additional storage system or network failures using brokering (e.g., using a cloud service). This combination of brokering and quorum can be further extended. For example, in a pod spread across four storage systems, any three can detach the fourth storage system, but if two in-sync storage systems are communicating with each other but not with the other two storage systems, they will currently consider themselves both in sync and will then use brokering to safely detach the other two. Even in a five-storage-system pod configuration, if four storage systems vote to detach the fifth storage system, then the remaining four can use mediation if they split into two equal halves; once the pod is down to two storage systems, they can use mediation to resolve successive failures. Five to three might use quorum among them, then allow a drop to two, with the remaining two storage systems again using mediation if there are additional failures. This general multi-mode quorum mediation mechanism can handle an additional number of situations that neither quorum nor independent mediation between symmetric storage systems can handle. This combination can increase the number of cases where a faulty or occasionally unreachable mediator can be reliably used (or in the case of cloud mediators where customers may not fully trust them). Furthermore, this combination better handles the case of a three-storage-system pod, where the first storage system may successfully detach the second and third storage systems during a network failure that only affects the first storage system, as a result of the mediation alone. This combination may also better handle a series of failures affecting one storage system at a time, as illustrated in the three-to-two-to-one example. These combinations work because the in-sync and detach operations create a specific state—in other words, moving from detached to in-sync is a process, and each stage in the quorum/mediator relationship series ensures that all online/in-sync storage systems agree on the current, durable state for the pod, so the system is stateful. This differs from some other cluster models, where simply having a majority of cluster nodes in communication again is expected to be sufficient to resume operation. However, a priority model may still be added, where satellite or top-of-rack storage systems never participate in mediation or quorum and only serve pods if they can connect to online storage systems that participate in mediation or quorum.

[0349] In some cases, the intermediation service (1500) or external pod membership manager may be located in a different failure zone than the failure zone for the synchronously replicated storage systems (1514, 1524). For example, when using two storage system pods (1501), if the two storage systems (1514, 1524) are separated into separate failure zones, for example, by physical location—one in a city and the other on the outskirts of the city, or one in a data center connected to one power grid or Internet access point and the other in another data center connected to a different power grid or Internet access point—then it is generally preferable to be in some other failure zone other than the two storage systems. As an example, the intermediation service (1500) may be in a different part of an extended urban area of a city or connected to a different power grid or Internet access point. However, the synchronously replicated storage systems may be in the same data center to provide better storage reliability, in which case network, power, and cooling zones may be taken into account.

15 includes requesting (1502) mediation by a first storage system (1514) from an intermediation service (1500) in response to detecting a trigger event. In this example, the trigger event may be a communications failure of a data communications link (1516) between the first storage system (1514) and the second storage system (1524), and detecting the failure may be due to a hardware failure, a failure to acknowledge receipt of a transmission, a failed retry effort, or some other method. In other cases, the trigger event may be the expiration of a synchronous replication lease, and requesting mediation may be part of an attempt to synchronize the connection and coordinate the resumption of an active lease. Such a lease may initially be established according to timing information for at least one of the multiple storage systems in a variety of different ways. For example, the storage systems may establish a synchronous replication lease by utilizing timing information to coordinate or exchange clocks for each of the multiple storage systems. In such an example, once the clocks are adjusted on each of the storage systems, the storage systems may establish synchronous replication leases that extend for a predetermined period beyond the adjusted or exchanged clock value. For example, if the clocks for each storage system are adjusted by time X, the storage systems may each be configured to establish synchronous replication leases that are valid for X+2 seconds. Additional description of adjusting or exchanging clocks may be found in U.S. Provisional Patent No. 62/518,071, which is incorporated herein by reference in its entirety.

[0351] Furthermore, requesting (1502) mediation from the intermediation service (1500) by the first storage system (1514) in response to detecting the trigger event may be implemented by a controller of the first storage system (1514) detecting the trigger event and sending a request (1560) to the intermediation service (1500) over the network (1554). In some examples, the intermediation service (1500) may be a third-party service that provides—for multiple computer systems—mutually exclusive access to resources, such as specific database entries for storing values. For example, the intermediation service (1500) may be provided by a database service provided by a cloud service provider, by a host computer issuing a request to modify a database, or by some third-party service providing mutually exclusive access to resources, and the resources may be storage, a state machine, or some other type of resource capable of indicating specific modifications based on a request from a particular client. In this example, after sending a request for mediation (1560), the first storage system (1514) waits (1503A) for an indication from the mediation service (1500) indicating a positive mediation result (1503B), a negative mediation result, or a lack of response (1503C). If the first storage system (1514) receives a negative mediation result or no mediation result (1503C), and if a threshold amount of time to wait has not been exceeded, the first storage system (1514) may continue to wait (1506) for more time. However, if the amount of time to wait exceeds the threshold amount, then the first storage system (1514) may determine that another computer system has won the mediation and may proceed by taking itself offline (1506). In some examples, as described above, a request for mediation may be received by the mediation service (1500) as an atomic compare-and-set operation that attempts to set the value of a shared resource (1552) that may also be the target of a compare-and-set operation received from another storage system of the storage systems that maintain the pod (1501), with the storage system that successfully sets the shared resource (1552) winning the mediation.

[0352] The example of FIG. 15 also includes the second storage system (1524) requesting mediation (1510) from the intermediation service (1500) in response to detecting the trigger event. The requesting mediation (1510) from the intermediation service (1500) in response to detecting the trigger event may be implemented similarly to the embodiment of requesting mediation (1502) from the first storage system (1514) in response to the trigger event. However, in this example, the second storage system (1524) may receive a failure message or some indication that the request for mediation (1562) was not successful in response to sending the request (1562) to the intermediation service—as opposed to the first storage system (1514) successfully mediating.

15 example method continues by having the first computer system (1514) process (1504)—on behalf of the second storage system (1524)—data storage requests directed to the dataset (1512) that is synchronously replicated across the first storage system (1514) and the second storage system (1524) in response to the positive brokered result indication (1564) from the brokering service (1500). Synchronous replication of the dataset (1512) implementing the pod (1501), in addition to receiving and processing data storage requests directed to the dataset (1512), may be performed as described with respect to Figures 8A and 8B of U.S. Provisional Patent Nos. 62/470,172 and 62/518,071, the entireties of which are incorporated herein by reference. In this example, as described above with respect to FIG. 15, in response to a positive mediation result indication (1564), the first storage system (1514) may be considered the winning storage system, and the first storage system (1514) may detach the storage system with which communication was lost. However, in other examples, mediation may be performed according to any of the other methods of mediation described or combinations of mediation methods.

[0354] In some examples, defining a preference for which storage system wins mediation among multiple storage systems that synchronously replicate dataset (1512) may be implemented by assigning a delay value to each of the multiple storage systems. For example, if a first storage system (1514) is designated as the preferred storage system, then the first storage system (1514) may be assigned a delay value of zero (0) before making a request for mediation from the mediation service. However, for non-preferred storage systems, the delay value may be assigned to be greater than zero, such as three seconds, or some other value that would generally result in the preferred storage system winning mediation simply due to communication loss between the synchronously replicated storage systems.

[0355] For additional explanation, FIG. 16 sets forth a flowchart illustrating an example method for mediating between storage systems that synchronously replicate data sets in accordance with some embodiments of the present disclosure. The example method illustrated in FIG. 16 is similar to the example method illustrated in FIG. 15 in that it also includes requesting mediation (1502) from a mediation service (1500) by a first storage system (1514) in response to detecting a trigger event, and processing (1504) a data storage request directed to a data set (1512) that is synchronously replicated across the first storage system (1514) and the second storage system (1524) by the first computer system (1514)—on behalf of the second storage system (1524)—in response to a positive mediation request indication (1564) from the mediation service (1500).

[0356] However, the example method shown in FIG. 16 further includes detaching (1602) a second storage system (1524) from the plurality of storage systems (1514, 1524) synchronously replicating the dataset (1512) in response to the indication (1564) of a positive annotation result. Detaching (1602) another storage system may be performed at a storage system receiving an indication of a positive mediation result from the mediation service (1500) by removing the no longer communicating storage system (1524) from the in-sync list of storage systems replicating the dataset (1512), such that the removal from the in-sync list causes the winning storage system (1514) to not attempt to synchronize the detached storage system for subsequently received requests to modify the dataset. In this example, there are two storage systems (1514, 1524), although other quantities of storage systems are contemplated in other examples.

[0357] For additional explanation, FIG. 17 sets forth a flowchart illustrating an example method for mediating between storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. The example method illustrated in FIG. 17 is similar to the example method illustrated in FIG. 15, as it also includes requesting mediation (1502) from a mediation service (1500) by a first storage system (1514) in response to detecting a trigger event, in this example the trigger event being a communication failure.

[0358] However, the example method shown in FIG. 17 differs from the example method shown in FIG. 15 because the example method shown in FIG. 17 does not include any activities or actions performed by the second storage system (1524). This difference allows for a storage system among the multiple storage systems to fail or otherwise become unresponsive, and for one or more other storage systems to require intermediation from the intermediary service (1500) to continue servicing data storage requests directed to the synchronously replicated data set (1512).

[0359] The example method shown in FIG. 17 includes detecting (1702) a failure in communication between a first storage system (1514) and a second storage system (1524), where the first storage system (1514) and the second storage system (1524) are included in a storage system that synchronously replicates a data set (1512). Detecting (1702) the communication failure may be performed as described above with respect to FIG. 15.

[0360] Additionally, the example method shown in FIG. 17 also includes, in response to the indication (1564) of a positive mediation result, detaching (1704) the second storage system (1524) from the plurality of storage systems (1514, 1524) that synchronously replicate the dataset (1512). Detaching (1704) the second storage system (1524) may be performed similarly to detaching (1602) the second storage system (1524) described with reference to FIG. 16.

[0361] For additional explanation, FIG. 18 sets forth a flowchart illustrating an example method for recovery of a storage system that synchronously replicates a dataset in accordance with some embodiments of the present disclosure. While the example method illustrated in FIG. 18 depicts an embodiment in which dataset (1812) is synchronously replicated across only two storage systems (1814, 1824, 1828), each of which may be independently coupled to one another via one or more data communication links (1816, 1818, 1820), the example illustrated in FIG. 18 can be extended to embodiments in which dataset (1812) is synchronously replicated across additional storage systems.

[0362] Multiple storage systems (1814, 1824, 1828) synchronously replicating a dataset (1842) may communicate with each other during normal operation to receive and process requests (1804) from the host (1802) computing device. However, in some instances, one or more of the storage systems (1814, 1824, 1828) may fail, reboot, upgrade, or otherwise become unavailable. Recovery in this situation is the process of bringing the synchronizing pod member storage systems together after a failure or some other service outage causes at least one of the synchronizing storage systems to suspend and possibly lose the context of in-flight operations. A "pod," as the term is used here and throughout the remainder of this application, may be embodied as a management entity representing a dataset, a collection of management objects and management operations, a collection of access operations to modify or read the dataset, and multiple storage systems. Such management operations may modify or query managed objects equally through any of the storage systems, with access operations to read or modify a dataset operating equally through any of the storage systems. Each storage system may store a separate copy of the dataset as an appropriate subset of the dataset stored and advertised for use by the storage system, and operations that modify managed objects or datasets performed and completed through any one storage system are reflected in subsequent managed objects to query the pod or subsequent access operations to read the dataset. Additional details regarding "pods" may be found in previously filed provisional patent application Ser. No. 62/518,071, which is incorporated herein by reference. In this example, only three storage systems (1814, 1824, 1828) are shown, but generally any number of storage systems may be part of the in-sync list synchronously replicating dataset (1812).

[0363] If any one or more storage systems that are members of a pod are interrupted, then any remaining storage systems, or any storage systems that resume operation early, may either detach them (so that they are no longer in sync) or wait for them before proceeding and engage in recovery actions to ensure consistency. If the outage is short enough and recovery is rapid enough, then operating systems and applications running outside the storage system or on a storage system that does not fault to stop the application itself, may experience a temporary delay in processing storage operations, but may not experience a service outage. SCSI and other storage protocols support retries involving an alternate target storage interface when operations are lost due to the outage of a primary storage controller or interface target controller; in particular, SCSI supports a BUSY status to request an initiator retry, which may be used while the storage controller participates in recovery.

[0364] Generally, one goal of recovery is to handle any inconsistencies from an unexpected disruption of ongoing distributed operations and to resolve the inconsistencies by making synchronized pod member storage systems sufficiently identical, at which point providing pod services can safely resume. Sufficiently identical includes at least the content stored in the pod; in other cases, sufficiently identical may include the state of persistent reservations. Sufficiently identical may also include ensuring that snapshots are either consistent—correct with respect to completed modification operations, concurrent modification operations, or more recently received modification operations—or consistently deleted. Depending on the implementation, there may be other metadata that should be made consistent. If there is metadata used to track or optimize the movement of content from a replication source to an asynchronous or snapshot-based replication target, then it may need to be made consistent to allow the replication source to switch smoothly from one member storage system of the pod to another. The existence and properties of volumes may also need to be restored, perhaps as well as definitions related to the application or initiating host system. Many of these properties may be recovered using standard database transaction recovery techniques, depending on how they are implemented.

[0365] In some cases, beyond ensuring that management metadata is sufficiently identical across storage systems that perform modifying operations on the contents of block-based storage systems, recovery must also ensure that those modifications are consistently applied or discarded across pods and with appropriate consideration of block storage semantics (ordering, concurrency, consistency, atomicity of operations such as COMPARE AND WRITE and XDWRITERREAD). Fundamentally, this embodiment relies on being able to know during recovery which operations may have been applied to at least one in-sync storage system for the pod that may not have been applied to all other in-sync storage systems for the pod, and then either apply them everywhere or undo them. Either action results in consistency—apply everywhere or undo everywhere—and there is no inherent reason why the answer must be uniform across all operations. Undo may be allowed if at least one in-sync storage system for the pod did not apply the operation. In general, it is often simpler to reason about applying all updates found on any in-sync storage system for a pod but not on all in-sync storage systems for the pod, rather than undoing some or all updates on one or more in-sync storage systems for the pod. To be efficient, knowing what was applied on some systems that may not have been applied on other systems generally requires the storage system to record something other than raw data (otherwise, all data may have to be compared, which would take an order of magnitude more time). Described below are additional details regarding implementations for recording such information that may enable storage system recovery.

[0366] Two examples of persistently tracking information to ensure consistency include (1) identifying that the contents of a volume may differ across synchronizing storage systems for a pod, and (2) identifying a collection of actions that may not have been universally applied across all synchronizing systems for a pod. The first example is the traditional model for mirroring: keeping a tracking map of logical areas that have been written (often as a list or as a bitmap that covers the logical space of the volume with some granularity), and using that list during recovery to note which areas may differ between one copy and another. The tracking map is written to some or all mirrors (or written separately) before or during the writing of volume data, such that recovery of the tracking map is guaranteed to cover any volume areas that were in flux at the time of the failure. Recovery in this first variant generally consists of copying the contents from one copy to another and verifying that they are the same.

[0367] A second example of persistence tracking—based on operation tracking—may be useful in storage systems that support synchronous replication of large volume range virtual copies within and between volumes within a pod, because this case may be more difficult or expensive to track simply as potential differences in volume content between synchronously replicated storage systems (but see the section describing tracking and recovery in content-addressable storage systems). Also, simple content tracking may not work as well in storage systems where synchronous replication must track more complex information, such as in a content tracking graph with extents and larger-granularity identifiers that drive a form of asynchronous replication, and where an asynchronous replication source is migrated or faulted over from one in-sync system to another storage system within a pod. When operations are tracked instead of content, recovery involves identifying operations that may not have completed anywhere. Once such operations are identified, any ordering consistency issues should be resolved, as they should be during normal runtime using techniques such as leader-defined ordering or predicates or through interlocked exceptions. Interlock exceptions are discussed below, and with regard to predicates, the description of the relationship between an operation and a common metadata update may be described as a set of interdependencies between separate modifying operations—these interdependencies may be described as a set of precursors on which an operation depends in some way, and the set of precursors may be viewed as predicates that must be true for the operation to complete. To continue with this example, given an identified operation, the operation may then be reapplied. Information recorded about the operation should include any metadata changes that must be consistent across pod member storage systems, and this recorded information may be copied and applied. Furthermore, predicates may not need to be protected if they are used to propagate restrictions on concurrency between leaders and followers, as the information persisted could imply diverse plausible conclusions if those predicates change the order in which storage systems persist information.

[0368] As described more thoroughly in U.S. Provisional Patent Application Nos. 62/470,172 and 62/518,071, the references of which are incorporated herein in their entireties, a collection of synchronizing storage systems may implement a symmetric I/O model to provide data consistency. In a symmetric I/O model, multiple storage systems may maintain datasets within a pod, and a member storage system receiving an I/O operation may process the I/O operation locally concurrently with the processing of the I/O operation on all other storage systems in the pod—the receiving storage system may initiate processing of the I/O operation on the other storage systems. However, in some cases, multiple storage systems may receive independent I/O operations that write to overlapping memory regions. For example, if a first write enters a first storage system, the first storage system may then begin persisting the first write locally while sending it to a second storage system; approximately simultaneously, a second write to a volume region that overlaps with the first write is received at the second storage system, which begins persisting the second write locally while sending the second write to the first storage system. In this situation, at some point, either the first storage system, the second storage system, or both storage systems may become aware of the simultaneous overlap. Furthermore, in this situation, the first write cannot complete at the first storage system until the second storage system has persisted the first write and responded with a success indication, and the first storage system has successfully persisted the first write; the second storage system is in a similar situation as the second write. Because both storage systems have access to both the first and second writes, either storage system may detect the concurrent conflict, and when one storage system detects the concurrent conflict, the storage system may trigger an exception, referred to herein as an "interlock exception." When the situation is expanded to additional storage systems, one solution involves two, or perhaps more, storage systems participating in an interlock exception to reach agreement on which write operation takes precedence.

[0369] In another example, such as in the case of overlapping write requests, write-type requests (e.g., WRITE requests, WRITE SAME requests, UNMAP requests, or a combination) that overlapped in time and volume address range at the time of the event that interrupted replication and ultimately led to its recovery may have completed inconsistently between synchronizing storage systems. How this situation is handled may depend on the implementation of the I/O path during normal operation. In this example, further below is described a first write and a second write that overlapped in time, each received by one storage system or another for the pod before either was signaled as completed. This example is easily extended to more than two writes by considering them two by two in sequence, and by considering that the first and second writes may have completed on multiple storage systems, and by considering that the first, second, and third writes (or additional writes) may have completed inconsistently on more than two storage systems. The described technique is easily extended to these cases. In a symmetric I/O-based storage system implementation based on interlock exceptions, it is possible that only the first write completed on one storage system. Meanwhile, the second of two overlapping writes may have completed on a second storage system. This case can be detected by noting that ranges overlap between each write and by noting that neither storage system contains an alternate overlapping write. If two writes completely overlap (one completely covers the other), then one of the two writes may simply be copied to the other storage system and applied to replace the contents of that storage system for that volume address range. If the writes only partially overlap, then the partially overlapping contents may be copied (applied) from one storage system to the other. Meanwhile, the non-overlapping portions can be copied between each storage system so that the contents are unified and up-to-date on both storage systems. In a leader-based system with predicates or some other means for the leader to declare that one write precedes another, the storage systems performing the writes may persist them overlappingly, or may persist both together. Alternatively, an implementation may persist the two writes separately and out of order, with the ordering predicate used solely to control completion signaling. If an implementation allows out-of-order write processing, then the preceding example explains how consistency can be recovered. If the storage systems enforce persistence ordering during normal operation, then recovery may only see the first write on the first storage system, excluding the first and second writes on the second storage system. In that case, the second write can be copied from the second storage system to the first storage system as part of recovery.

[0370] In another example, snapshots may also be recovered. In some cases, such as in the case of snapshots that coincide with a modification, where some modifications of the leader decision should be included in the snapshot and others should not, the recorded information may include information about whether a particular write should be included in the snapshot. In that model, it may not be necessary to ensure that everything the leader decided to include in the snapshot must eventually be included in the recovered snapshot. If one synchronizing storage system for a pod recorded the existence of a snapshot, and no synchronizing storage system for the pod recorded a write that was ordered for inclusion in the snapshot, then uniformly applying the snapshot without including that write would result in snapshot content that is completely consistent across all synchronizing storage systems for the pod. This inconsistency should only occur in the case of concurrent writes and snapshots that were never signaled as completed, and therefore containment guarantees are not guaranteed. That is, leader assignment and ordering of predicates may be required only for runtime consistency rather than for recovery ordering consistency. If the recovery identifier identifies the write for inclusion in the snapshot, but the recovery does not position the write, the snapshot operation itself may safely ignore the snapshot, depending on the implementation. The same arguments for snapshots also apply to virtual copies of volume address ranges via SCSI EXTENDED COPY and similar operations. That is, the reader defines which writes to the source address range may logically precede the copy, and which writes to the target address range may logically precede or follow the address range copy. However, during recovery, the same arguments apply as for snapshots. That is, a write that occurs concurrently with a volume range copy may miss either the concurrent write or the volume range copy, provided the results are consistent across all in-sync storage systems for the pod, do not revert modifications completed anywhere, and do not reverse modifications that dataset consumers may have read.

[0371] Further with regard to this example illustrating snapshot recovery, if any storage system applied the write for a COMPARE AND WRITE, then the comparison must have succeeded at one in-sync storage system for the pod, and since runtime consistency meant that the comparison should have succeeded at all in-sync storage systems for the pod, if any such storage system applied the write, it may be copied and applied to any other in-sync storage systems for the pod that had not applied it before recovery. Furthermore, recovery of an XDWRITERREAD or XPWRITE request (or similar arithmetic transformation operation between existing data and new data) may work by delivering the results of the transformation after the resulting read from one storage system, or it may work by delivering the transforming write along with the data to other storage systems, if it can be assured that any ordering data preceding the transforming write is consistent across in-sync storage systems for the pod, and if it can be reliably determined which such storage systems have not yet applied the transforming write.

[0372] As another example, metadata recovery may be performed. In this case, recovery should also result in consistent recovery of metadata between synchronized storage systems for the pod, and the metadata is expected to be consistent across the pod. If this metadata is included with operations, it can be applied along with the content updates described by those operations. How this data is merged with existing metadata depends on the metadata and implementation. Longer-term change tracking information for driving asynchronous replication can often be merged very simply as neighboring or otherwise related modifications are identified.

[0373] As another example, recording recent activity for operation tracking may be implemented in a variety of ways to identify operations that were in progress at synchronized storage systems in a pod at the time of a failure or other type of service interruption that led to recovery. For example, one model is to record recovery information for modifications to each synchronized storage system in a pod, either atomically with any modification (which can work well if updates are staged through a high-speed journal device) or by recording information about operations that were in progress before the operation occurred. The recorded recovery information may include a logical operation identifier, such as based on the original request or some identifier assigned by the reader as part of describing the operation; any level of operation description may be required for recovery to operate. Information recorded by the storage system for writes included in the contents of a concurrent snapshot should indicate that the write is included in the snapshot, not just in the contents of the volume to which the write is applied. In some storage system implementations, the contents of the snapshot are automatically included in the contents of the volume unless they are replaced by specific duplicate content in a newer snapshot or by specific duplicate content later written to the volume live. Two concurrent write-type requests (e.g., WRITE, WRITE SAME, or UNMAP requests, or a combination) that overlap in time and in volume address may be explicitly ordered by the leader to ensure that the first write is first persisted to all synchronizing storage systems for the pod before the second write can be persisted by any synchronizing storage system for the pod. This simply ensures that inconsistencies do not occur. Furthermore, this may be acceptable because concurrent overlapping writes to a volume are very rare. In that case, if there is a record on any recovering storage system for the second write, then the first write must have completed everywhere so that it should not require recovery. Alternatively, a predicate may be written by the leader requiring that the storage system order the first write before the second write. The storage system may then execute both writes together, thereby ensuring that they either both persist or neither fail to persist. In other cases, the storage system may persist the first write, then the second, after the durability of the first write is guaranteed. COMPARE AND WRITE, XDWRITERREAD, or XPWRITE requests should be ordered so that the precursor content is identical on all storage systems when each performs the operation. Alternatively, one storage system may compute the results and deliver the request to all storage systems as a normal write-type request. Furthermore, with regard to making these operations recoverable, tracking which operations have completed everywhere may allow their recency to be downplayed; recorded information that would otherwise cause operation recovery analysis for completed operations may then be either discarded by recovery or effectively ignored.

[0374] In another example, discarding completed operations may be implemented. One example for handling the clearing of recorded information is to clear an operation across all synchronized storage systems for a pod once it is known to have been processed across all synchronized storage systems. This can be implemented by having the storage system that received the request and signaled completion send a message to all synchronized storage systems for the pod after completion is signaled, allowing each storage system to discard them. Recovery then involves querying all synchronized storage systems for the pod involved in the recovery for all recorded operations that have not been discarded. Instead, these messages could be batched so that they occur periodically (e.g., every 50 ms) or after some number of operations (e.g., every 10 to 100). This batching process could significantly reduce message traffic at the expense of somewhat increased recovery times, as more fully completed operations would potentially be reported as incomplete. Furthermore, in a leader-based implementation (by way of example), the leader would be made aware of which operations have completed, and the leader could send out a clear message.

[0375] In another example, a sliding window may be implemented. Such an example may work well in a leader-and-follower based implementation, where the leader may assign sequence numbers to operations or collections of operations. Thus, in response to the leader determining that all operations up to a certain sequence number are complete, the leader may send a message to all in-sync storage systems indicating that all operations up to that sequence number are complete. Sequence numbers can also be arbitrary numbers, whereby when all operations associated with a given number are completed, a message is sent indicating that all those operations are complete. Using a sequence number-based model, recovery would query any in-sync storage system for all operations associated with a sequence number greater than the last completed sequence number. In a leaderless, symmetric implementation, each storage system receiving requests for a pod could define its own sliding window and sliding window identity space. In that case, recovery may include querying for all operations in any in-sync storage window whose sliding window identity is associated with any sliding window identity space that is after the last completed identity, and for which operations for all preceding identifiers have also completed.

[0376] In another example, checkpoints may be implemented. In a checkpoint model, a special operation may be inserted by the leader that depends on the completion of a uniform set of pre-cursor operations, upon which all subsequent operations then depend. Each storage system may then persist the checkpoint in response to all pre-cursor operations being persisted or completed. A subsequent checkpoint may be initiated some time after the previous checkpoint has been signaled as persisted at all in-sync storage systems for the pod. A subsequent checkpoint would not be initiated in this manner until some time after all pre-cursor operations have been persisted across the pod. Otherwise, the previous checkpoint would not be completed. In this model, recovery may involve querying all in-sync storage systems for all operations following the previous to last checkpoint. This would be accomplished by identifying the second through last checkpoints known to any in-sync storage systems for the pod, or by asking each storage system to report all operations since the second through last persisted checkpoint. Alternatively, recovery may involve retrieving the last checkpoint known to have completed on all in-sync storage systems, and querying all subsequent operations on any in-sync storage systems - if a checkpoint has completed on all in-sync storage systems, then all operations prior to that checkpoint have been specifically persisted everywhere.

[0377] In another example, pod recovery may be performed based on a replicated directed acyclic graph of logical extents. However, before describing such an embodiment, a storage system using a directed acyclic graph of logical extents will first be described.

[0378] A storage system may be implemented based on a directed acyclic graph containing logical extents. In this model, logical extents may be categorized into two types: leaf logical extents, which refer to some amount of data stored in some way, and composite logical extents, which refer to other leaf or composite logical extents.

[0379] Leaf extents can reference data in a variety of ways. A leaf extent may directly point to a single range of stored data (e.g., 64 kilobytes of data), or it may be a collection of references to stored data (e.g., a 1 megabyte "range" of contents that maps some number of virtual blocks associated with a range to physically stored blocks). In the latter case, these blocks may be referenced using some identity, and some blocks within the extent's range may not map to anything. Also, in the latter case, these block references need not be unique, allowing multiple mappings from virtual blocks in some number of logical extents within and across some number of volumes to map to the same physically stored block. Instead of stored block references, logical extents could encode simple patterns. For example, blocks that are strings of identical bytes could simply encode that the block is a repeated pattern of identical bytes.

[0380] A composite logical extent may be a logical range of content having some virtual size that includes multiple maps, each mapping a logical range of content from a subrange of the composite logical extent to an underlying leaf logical extent or composite logical extent. Transforming a request related to content in a composite logical extent then involves taking the content range for the request within the context of the composite logical extent, determining which underlying leaf logical extent or composite logical extent the request maps to, and transforming the request to apply to the appropriate range of content within those underlying leaf logical extents or composite logical extents.

[0381] Volumes, i.e., files or other types of storage objects, can be described as composite logical extents. Thus, these represented storage objects (which will be referred to simply as volumes in most of the description) can be organized using this extent model.

[0382] Depending on the implementation, a leaf logical extent or a composite logical extent may be referenced from multiple other composite logical extents, essentially enabling inexpensive duplication of larger collections of content within and across volumes. Thus, logical extents may essentially be arranged in an acyclic graph of references, each terminating in a leaf logical extent. This can be used to create copies of volumes, to create snapshots of volumes, or as part of virtual range copy support within and between volumes as part of an EXTENDED COPY or similar type of operation.

[0383] Implementations may provide each logical extent with an identity that can be used to name it. This simplifies referencing, as the reference within a composite logical extent will be a list containing the logical extent identifier and the logical subrange corresponding to each such logical extent identity. Also, within a logical extent, each stored data block reference may be based on some identity used to name it.

[0384] To support these overlapping uses of extents, an additional feature can be added: copy-on-write logical extents. When a modification operation affects a copy-on-write leaf logical extent or composite logical extent, the logical extent is copied, and the copy is a new reference and possibly has a new identity (depending on the implementation). The copy preserves all references or identities related to the underlying leaf logical extent or composite logical extent, regardless of any modifications resulting from the modification operation. For example, a WRITE, WRITE SAME, XDWRITERREAD, XPWRITE, or COMPARE AND WRITE request may store new blocks in the storage system (or use deduplication techniques to identify existing stored blocks) and modify the corresponding leaf logical extent to reference or store identities for the new set of blocks, possibly replacing references and stored identities for the previous set of blocks. Alternatively, an UNMAP request may modify the leaf logical extent to remove one or more block references. In both types of cases, a leaf logical extent is modified. If the leaf logical extent is copy-on-write, then a new leaf logical extent is created, formed by copying unaffected block references from the old extent and then replacing or deleting block references based on the modification operation.

[0385] The composite logical extent used to locate the leaf logical extent may then be modified to store a new leaf logical extent reference or identity associated with the copied and modified leaf logical extent as a replacement for the previous leaf logical extent. If the composite logical extent is copy-on-write, then a new composite logical extent is created with the new reference or identity, any unaffected references or identities to its underlying logical extents are copied to the new composite logical extent, and the previous leaf logical extent reference or identity is replaced with the new logical extent reference or identity.

[0386] This process continues from the referenced extent back to the referencing composite extents, based on a search path through the acyclic graph used to process the modification operation, until all copy-on-write logical extents are copied, modified, and replaced.

[0387] These copied leaf logical extents and composite logical extents may then exclude the copy-on-write feature, whereby additional modifications do not result in additional copies. For example, the first time some underlying logical extent in a copy-on-write "parent" composite extent is modified, that underlying logical extent is copied and modified, and the copy then has a new identity written into the copied, replaced instance of the parent composite logical extent. However, the second time some other underlying logical extent is copied and modified, a new identity for the other underlying logical extent copy is written into the parent composite logical extent, and the parent may then be modified in place without additional copies, replacing the necessary references to the parent composite logical extent.

[0388] Modification operations to new areas of a volume or composite logical extent for which there is no current leaf logical extent may create a new leaf logical extent to store the results of those modifications. If that new logical extent is referenced from an existing copy-on-write logical extent, then that existing copy-on-write composite logical extent references the new logical content, resulting in another sequence of copy, modify, and replace operations similar to the sequence for modifying an existing leaf logical extent.

[0389] If a parent composite logical extent cannot grow large enough (based on the implementation) to cover the associated address range, including the new leaf logical extent to create for the new modification operation, then the parent composite logical extent may be copied to two or more new composite logical extents that are in turn referenced from a single "grandchild" composite logical extent that is itself a new reference or new identity. If that grandchild logical extent is found through another composite logical extent that is itself copy-on-write, then that other composite logical extent is copied, modified, and replaced in a manner similar to that described in the preceding paragraph. This copy-on-write model can be used as part of implementing snapshots, volume copies, and virtual volume address range copies in storage system implementations based on these directed acyclic graphs of logical extents. To create a snapshot as a read-only copy of an otherwise writable volume, the graph of logical extents associated with the volume is marked copy-on-write, and a reference to the original composite logical extent is maintained by the snapshot. Modification operations on the volume then create logical extent copies as needed, resulting in a volume that stores the results of those modification operations and a snapshot that preserves the original contents. Volume copies are similar, except that both the original and copied volumes modify content, resulting in their own copied logical extent graphs and subgraphs.

[0390] Virtual volume address range copies may operate by copying block references within and between leaf logical extents (which does not itself involve using copy-on-write techniques unless changes to the block references modify the copy-on-write leaf logical extents). Alternatively, virtual volume address range copies may duplicate references to leaf logical extents or composite logical extents, which works well for volume address range copies of larger address ranges. Furthermore, this allows the graph to be a directed acyclic graph of references rather than simply a reference tree. Copy-on-write techniques associated with duplicated logical extent references can be used to ensure that modification operations on the source or target of a virtual address range copy result in the creation of new logical extents to store those modifications without affecting the target or source that share the same logical extent immediately after the volume address range copy operation.

[0391] Input/output operations for a pod may also be performed based on replicating a directed acyclic graph of logical extents. For example, each storage system in a pod may implement a private graph of logical extents, such that the graph on one storage system for the pod has no particular relationship to the graph on any second storage system for the pod. However, there is value in synchronizing the graphs between storage systems within a pod. This may be useful for resynchronization and for coordinating features such as asynchronous replication or snapshot-based replication to remote storage systems. Furthermore, it may also be useful for handling snapshot distribution and reducing some overhead for copy relationship processing. In such a model, keeping the contents of a pod in sync across all synchronizing storage systems for the pod is essentially the same as keeping the graphs of leaf logical extents and composite logical extents in sync for all volumes across all synchronizing storage systems for the pod and ensuring that the contents of all logical extents are in sync. To be in sync, matching leaf logical extents and composite logical extents should either have the same identity or should have mappable identities. Mapping might involve some set of intermediate mapping tables, or it might involve some other type of identity translation. In some cases, the identities of blocks mapped by leaf logical extents could also be kept in sync.

[0392] In a leader-and-follower based pod implementation with a single leader per pod, the leader may be responsible for determining any changes to the logical extent graph. When a new leaf logical extent or composite logical extent is created, it may be given an identity. When an existing leaf logical extent or composite logical extent is copied to form a new logical extent with modifications, the new logical extent may be described as a copy of the previous logical extent with some set of modifications. When an existing logical extent is split, the split may be described along with the resulting new identity. When a logical extent is referenced as a base logical extent from some additional composite logical extent, the reference may be described as a change to the composite logical extent to reference that base logical extent.

[0393] Modification operations in a pod thus include distributing descriptions of modifications to the logical extent graph (when new logical extents are created to extend content, or when logical extents are copied, modified, or replaced to handle copy-on-write conditions related to snapshots, volume copies, and volume address range copies), and distributing descriptions and contents for modifications to the content of leaf logical extents. An additional advantage of using metadata in the form of a directed acyclic graph, as described above, is that I/O operations that modify stored data in physical storage can be effected at the user level—without modifying the data stored in physical storage—by modifying the metadata corresponding to the data stored in physical storage. In disclosed embodiments of a storage system in which the physical storage may be a solid-state drive, wear associated with modifications to flash memory can be avoided or reduced due to I/O operations being effected by modifying metadata representing the data targeted by the I/O operation, instead of through a read, erase, or write to flash memory. Furthermore, in a virtualized storage system, the metadata described above may be used to handle the relationship between virtual addresses, i.e., logical addresses, and physical addresses, i.e., real addresses - in other words, the metadata representation of stored data enables a virtualized storage system that may be considered flash-friendly in that it reduces or minimizes wear on flash memory.

[0394] The leader storage system may perform its own local operations to implement these descriptions in the context of its local copy of the pod dataset and the metadata of its local storage system. Additionally, each synchronizing follower performs its own separate local operations to implement these descriptions in the context of their separate local copies of the pod dataset and the metadata of their separate local storage systems. When both the leader and follower are complete, the result is a compatible graph of logical extents with compatible leaf logical extent contents. These graphs of logical extents then become a kind of "common metadata," as described in the example above. This common metadata may be described as dependencies between modification operations and required common metadata. Transformations to the graph may be described as separate operations with queued predicate relationships with subsequent modification operations. Instead, each modification operation may rely on the same particular graph transformation that is not yet known to complete across the entire pod. Each modification operation may include portions of any graph transformations it relies on. Processing an operation description that identifies a "new" leaf or composite logical extent that already exists may avoid creating a new logical extent because that portion was already processed in the processing of some earlier operation; instead, it may implement only the portion of the operation that modifies the contents of the leaf or composite logical extent. It is the reader's responsibility to ensure that the transformations are compatible with each other. For example, a pod may start with two writes receiving the pod. The first write replaces composite logical extent A with a copy created as composite logical extent B, replaces leaf logical extent C with a copy as leaf logical extent D and modifications to store the content for the second write, and writes leaf logical extent D to composite logical extent B. Meanwhile, the second write implies the same copy and replacement of composite logical extent A with composite logical extent B, but copies a different leaf logical extent E and replaces it with logical extent F that is modified to store the content of the second write, and writes logical extent F to logical content B. In that case, the description for the first write might include replacing A with B and C with D, and writing D to composite logical extent B and writing the contents of the first write to leaf extent B, and the description for the second write might include replacing A with B and E with F, and writing F to composite logical extent B, with the contents of the second write written to leaf extent F. The leader or any follower may then process the first write or the second write in any order, with the end result being B copying and replacing A, D copying and replacing C, F copying and replacing E, and D and F being written to composite logical extent B. The second copy of A to form B can be avoided by recognizing that B already exists. In this way, the leader may ensure that compatible common metadata is maintained for the logical extent graph across synchronizing storage systems for the pod.

[0395] Given an implementation of a storage system that uses a directed acyclic graph of logical extents, recovery of a pod may be performed based on the replicated directed acyclic graph of logical extents. Specifically, in this example, recovery at a pod may be based on the replicated extent graph and then involve not only recovering the contents of leaf logical extents but also restoring consistency to these graphs. In this implementation of recovery, operations may include querying not only for all leaf logical extent content modifications that are not known to be completed across all storage systems for the pod, but also for graph transformations that are not known to be completed across all synchronizing storage systems for the pod. Such queries may be based on operations since some coordinated checkpoint, or simply operations that are not known to be completed; each storage system maintains a list of operations during normal operation that have not yet signaled as completed. In this example, the graph transformations are simple. Graph transformations create new things, copy old things to new things, and may copy old things to two or more separate new things; they modify composite extents to modify their references to other extents. Any stored operation description found on any in-sync storage system that creates or replaces any logical extent may be copied and executed on any other storage system that does not already have that logical extent. As long as the involved leaf logical extent or composite logical extent has been properly recovered, operations that describe modifications to a leaf logical extent or composite logical extent can apply those modifications to any in-sync storage system that has not yet applied them.

[0396] Further in this example, pod recovery may include:
creating leaf logical extents and composite logical extents that were not known to be completed on all in-sync storage systems for the pod, and querying all in-sync storage systems for their precursor leaf logical extents and composite logical extents, if any;
Querying all in-sync storage systems for the pod for corrective actions on leaf logical extents that were not known to be completed on all in-sync storage systems;
Querying logical address range copy operations as new references to existing leaf logical extents and compound logical extents;
Identifying modifications to a leaf logical extent that are not known to be completed and that are the source for an alternate leaf logical extent that may also need to be recovered, so that modifications can be made to that leaf logical extent on all in-sync storage systems before the leaf logical extent copy is recovered on any in-sync storage systems that did not already copy it;
Completing the copy operations of all leaf logical extents and compound logical extents;
Applying all additional updates to leaf logical extents and compound logical extents, including naming new logical extent references, updating leaf logical extent contents, or deleting logical extent references;
- May include determining that all necessary actions have been completed, at which point additional aspects of recovery can proceed.

[0397] In another example, as an alternative to using a logical extent graph, storage may be implemented based on a replicated content-addressable store. In a content-addressable store, for each block of data (e.g., every 512 bytes, 4096 bytes, 8192 bytes, or even every 16384 bytes), a unique hash value (sometimes called a fingerprint) is calculated based on the block contents, allowing a volume or volume extent range to be described as a list of references to blocks with particular hash values. In a synchronously replicated storage system implementation based on references to blocks with the same hash value, replication might involve a first storage system receiving blocks, calculating fingerprints for those blocks, identifying block references for those fingerprints, and delivering the changes to one or more additional storage systems as updates to the mapping of volume blocks to the referenced blocks. If a block is found to be already stored by the first storage system, that storage system can use that reference to name the reference on each of the additional storage systems (either because the references use the same hash value, or because the identifiers for the references are identical or easily mappable). Alternatively, if a block cannot be found by a first storage system, then the contents of the first storage system, along with the hash value or identity associated with the block contents, may be delivered to the other storage system as part of the operation description. Furthermore, the volume description of each synchronizing storage system is then updated with the new block reference. Recovery in such a store may then include comparing recently updated block references for the volume. If block references differ between different synchronizing storage systems for a pod, then one version of each reference may be copied to the other storage system to make them consistent. If a block reference on one system does not exist, then it is copied from some storage system that stores the block for that reference. Virtual copy operations may be supported in such block or hash reference stores by copying the reference as part of performing the virtual copy operation.

[0398] For a particular embodiment for system recovery, the example method shown in FIG. 18 includes receiving (1842) a request (1804) to modify dataset (1812) by at least one storage system among a plurality of storage systems (1814, 1824, 1828) that synchronously replicate dataset (1812). Receiving (1842) the request (1804) to modify dataset (1812) may be implemented similarly to receiving (1806) the request (1804) to modify dataset (1812).

[0399] The example method shown in FIG. 18 also includes generating (1844) recovery information (1852) indicating whether the request (1804) to modify the dataset (1812) has been applied at all storage systems of the multiple storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812). Generating (1844) recovery information (1852) indicating whether the request (1804) to modify the dataset (1812) has been applied at all storage systems (1814, 1824, 1828) of the plurality of storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812) may be implemented using various techniques described above, including differential tracking-based recovery, operation tracking-based recovery including duplicate write recovery, snapshot recovery, metadata and common metadata recovery, recovery based on recording recent activity for operation tracking including discarding completed operations, using a sliding window, and using checkpoints, pod recovery based on a replicated directed acyclic graph of logical extents, and tracking and recovery in a replicated content-addressable store. In short, various techniques may be used to generate recovery information, and the recovery information indicates which storage systems (1814, 1824, 1828) of the plurality of storage systems (1814, 1824, 1828) the request (1804) to modify the dataset (1812) applied to.

18 also includes applying (1846) recovery actions in response to the system failure according to recovery information (1852) indicating whether the request to modify has been applied at all storage systems of the plurality of storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812). The recovery actions may be performed by applying the request (1804) to modify the dataset (1812) at all storage systems that did not apply the request (1804) to modify the dataset (1812)—the recovery information (1852) may include tracking information indicating which storage systems among the plurality of storage systems (1814, 1824, 1828) applied or did not apply one or more requests to modify the synchronously replicated dataset (1812), including the most recently received request (1804). However, in other cases, recovery action may be performed by canceling or undoing the application of request 1804 to modify dataset 1812 on a collection of storage systems that have completed or partially completed application of request 1804. Generally, the default recovery action may be to identify each storage system that did not successfully complete request 1804 and apply request 1804 to it in addition to other outstanding requests that modify dataset 1812. Other embodiments of recovery action are described above with respect to recovery based on difference tracking, recovery based on operation tracking including recovery of duplicate writes, snapshot recovery, recovery of metadata and common metadata, recovery based on recording recent activity for operation tracking including discarding completed operations, using a sliding window, and using checkpoints, recovery of pods based on a replicated directed acyclic graph of logical extents, and tracking and recovery in a replicated content-addressable store.

[0401] For further explanation, Figure 19 illustrates a flowchart showing an example method of recovery for a storage system that synchronously replicates a dataset in accordance with some embodiments of the present disclosure. The example method illustrated in FIG. 19 is similar to the example method illustrated in FIG. 18 in that it also includes receiving (1842) a request to modify a dataset (1812) by at least one storage system among a plurality of storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812); generating (1844) recovery information (1852) indicating whether the request to modify the dataset (1804) has been applied at all storage systems of the plurality of storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812); and applying (1846) a recovery action in accordance with the recovery information (1852) indicating whether the request to modify has been applied at all storage systems of the plurality of storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812).

[0402] However, the example method shown in FIG. 19 further specifies that generating recovery information (1844) includes querying other storage systems of the plurality of storage systems for operations that have been confirmed as processed (1902), determining a set of storage systems for which operations have not been confirmed as completed (1904), and applying recovery actions (1846) includes completing the operations that have not been confirmed as completed for the set of storage systems (1906).

[0403] Querying other storage systems of the plurality of storage systems for operations that have been confirmed as completed or processed (1902) may be performed as described above with respect to operation tracking for a storage system that supports synchronous replication of virtual copies of volume ranges within and between volumes within a pod. Specifically, as described above with respect to discarding completed operations across all synchronizing storage systems for a pod after the operations have been confirmed as completed across all synchronizing storage systems for the pod, this may be performed by having the storage system that received the request and signaled completion send messages to all storage systems for the pod after completion has been signaled, allowing each storage system to discard them. Recovery then involves querying for all recorded operations that have not been discarded across all synchronizing storage systems for the pod involved in the recovery.

[0404] Determining (1904) the set of storage systems for which operations have not been confirmed as completed may be performed based on results from querying (1902) other storage systems, the set of storage systems being populated by one or more storage systems for which querying (1902) included a list of operations that had not been discarded.

[0405] Completing operations that have not been confirmed as completed for the collection of storage systems (1906) may be performed by reissuing the operations to the collection of storage systems, and for each uncompleted operation, sending information describing modifications to the dataset according to the corresponding request (1812), as described herein, and completing the steps described herein.

[0406] For further explanation, Figure 20 illustrates a flowchart showing an example method of recovery for a storage system that synchronously replicates a dataset in accordance with some embodiments of the present disclosure. The example method illustrated in FIG. 20 is similar to the example method illustrated in FIG. 18 in that it also includes receiving (1842) a request to modify a dataset (1812) by at least one storage system among a plurality of storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812); generating (1844) recovery information (1852) indicating whether the request to modify the dataset (1812) has been applied at all storage systems among the plurality of storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812); and applying (1846) a recovery action in accordance with the recovery information (1852) indicating whether the request to modify has been applied at all storage systems among the plurality of storage systems (1814, 1824, 1828) that synchronously replicate the dataset (1812).

[0407] However, the example method shown in FIG. 20 further specifies that generating recovery information (1844) includes generating recovery information (1852) (2002) by tracking progress toward applying requests (2004) to modify datasets (1812) at multiple storage systems.

[0408] Generating (2002) recovery information (1852) by tracking progress toward applying request (1804) to modify dataset (1812) at multiple storage systems may be performed as described above by using checkpointing to determine operations that are confirmed as processed or completed. In this manner, generated (2002) recovery information (1852) may indicate which storage systems have not processed or completed request (1804) to modify dataset (1812).

[0409] Applying (2004) the request (1804) to modify the dataset (1812) may be performed by using the recovery information (1852) to identify one or more storage systems for reissuing the request (1804), which may be performed as described above, sending information describing the modifications to the dataset in accordance with the request (1804) for the request (1804) to modify the dataset (1812), and completing the steps described above.

[0410] Undoing (2006) a request (1804) to modify dataset (1812) at a storage system that did not apply the request to modify dataset (1812) may be performed by using recovery information (1852) to identify one or more storage systems on which the request (1804) was processed or completed. Furthermore, undoing (2006) the request may rely on maintaining, for each storage system on which the request (1804) was completed, a log of changes corresponding to each request to modify dataset (1812), where each request to modify dataset (1812) may be further associated with an identifier. The log may also associate, for each request identifier, a version of a metadata representation. The version of the metadata representation may include a directed acyclic graph representing the state of the dataset prior to applying the request identifier. In some examples, such versioning information may correspond to a snapshot. As described above, given the visualization representation of the dataset, and given that only differences to the metadata representation of the dataset corresponding to a particular request are stored in addition to data overwritten by the corresponding request to modify the dataset, the storage requirements of the log should be minimized. In this manner, the storage system's controller may use the log to restore the dataset's correspondence to its prior state before application of request (1804) and define the current state of the metadata representation to its prior state before application of request (1804).

[0411] For additional explanation, FIG. 21 sets forth a flowchart illustrating an example method of resynchronization of storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (2114, 2124, 2128) illustrated in FIG. 21 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage systems (2114, 2124, 2128) illustrated in FIG. 21 may include the same components as the storage systems described above, fewer components, or additional components.

[0412] The example storage system configuration shown in FIG. 21 includes a dataset (2112) and multiple storage systems (2114, 2124, 2128) across which dataset (2112) may be synchronously replicated. The dataset (2112) shown in FIG. 21 may be embodied, for example, as the contents of a particular volume, as the contents of a particular shard of a volume, or as any other collection of one or more data elements. The dataset (2112) may be synchronized across the multiple storage systems (2114, 2124, 2128) such that each storage system (2114, 2124, 2128) maintains a local copy of dataset (2112). In the example described herein, such dataset (2112) is synchronously replicated across storage systems (2114, 2124, 2128) such that dataset (2112) can be accessed through any of storage systems (2114, 2124, 2128) that have performance characteristics such that any one storage system in the cluster will operate substantially less optimally than any other storage system in the cluster, at least as long as the cluster and the particular storage system being accessed are nominally performing. In such a system, modifications to dataset (2112) should be made to copies of the dataset resident on each storage system (2114, 2124, 2128) so that accessing dataset (2112) on any of storage systems (2114, 2124, 2128) produces consistent results. For example, a write request issued to a dataset must be serviced by all storage systems (2114, 2124, 2128) or by none of the storage systems (2114, 2124, 2128).

[0413] Additionally, for a dataset (2112), the multiple storage systems (2114, 2124, 2128) across which the dataset (2112) is synchronously replicated may be implemented, for example, by examining a pod definition or similar data structure that associates the dataset (2112) with one or more storage systems (2114, 2124, 2128) that nominally store the dataset (2112). In such an example, the pod definition may include at least one identification of the dataset (2112) and a collection of storage systems (2114, 2124, 2128) across which the dataset (2112) is synchronously replicated. Such a pod may encapsulate several (possibly optional) pieces of data consistency, including symmetric access, flexible addition/removal of replicas, highly available data consistency, uniform user management across storage systems in relation to the dataset, administrative host access, application clustering, etc. A storage system can be added to a pod, and the pod's dataset (2112) is copied to that storage system and then kept up to date as the dataset (2112) is modified. A storage system can also be removed from a pod, and the dataset (2112) is no longer kept up to date with the removed storage system. In such an example, the pod definition or similar data structure may be updated as storage systems are added to and removed from particular pods.

[0414] The example storage system shown in FIG. 21 also includes one or more data communication links (2116, 2118, 2120) between each of the multiple storage systems (2114, 2124, 2128) used to synchronously replicate the data set (2112). In the example method shown in FIG. 21, the storage systems (2114, 2124, 2128) within a pod communicate with each other for both high-bandwidth data transfers, as well as cluster communication, status communication, and management communication. These separate types of communication may be over the same data communication links (2116, 2118, 2120), or in alternative embodiments, these separate types of communication may be over separate data communication links (2116, 2118, 2120).

[0415] Additional details for implementing a storage system that synchronously replicates data sets may be found in U.S. Provisional Patent Nos. 62/470,172 and 62/518,071, which are incorporated by reference in their entireties.

[0416] As shown in FIG. 21, multiple storage systems (2114, 2124, 2128) synchronously replicate a dataset (2112) and may communicate with each other during normal operation to receive and process I/O requests from host computing devices. However, in some instances, one or more of the storage systems (2114, 2124, 2128) may fail, reboot, upgrade, or otherwise become unavailable, resulting in one or more of the storage systems (2114, 2124, 2128) becoming out of sync. To resume normal operation, the in-sync and out-of-sync storage systems undergo recovery operations and resynchronization—recovery is described in more detail in Application Serial No. 15/696,418, which is incorporated herein by reference in its entirety, and resynchronization is described below.

[0417] Initial synchronization of a storage system being added to a pod—or subsequent resynchronization of a storage system detached from a pod—involves copying all content, or all missing content, from a synchronizing storage system for the pod to an uninitialized, or unsynchronized, storage system before that storage system is brought online for active use in providing pod services. Such an initial synchronization may be performed for each storage system being introduced as an extension of the pod.

[0418] The difference between an initial synchronization of content to a storage system added to a pod and resynchronizing a storage system that has become out of sync through some set of events with the in-sync storage system for the pod is conceptually very similar. For example, in the case of a resynchronization, all blocks that may differ between the in-sync pod member storage system and the out-of-sync pod member are brought up to date before the out-of-sync pod member can come back online as the in-sync pod member storage system for the pod. In an initial synchronization, this may involve updating all blocks, and as a result, it is conceptually similar to a resynchronization where all blocks may differ. In other words, an initial synchronization may be considered equivalent to reattaching a storage system that was detached at the beginning of the pod, before any volumes are modified from their initial state or before any volumes are created or added to the pod.

[0419] Generally, resynchronization accomplishes at least two things to return a detached pod to a point where the pod is in sync and can be brought back online: (a) undoing, overwriting, or otherwise replacing any changes persisted in the detached pod when it was detached and that were not being held by an in-sync pod member, and (b) updating the attaching storage system to match the content and common metadata for the pod. To be brought back online, reattaching a storage system may include reenabling synchronous replication, reenabling symmetric synchronous replication, and reenabling the receiving and processing of operations for the pod at the reattached storage system. Operations for the pod may include reads, data modification operations, or management operations.

[0420] In the process of detaching a storage system, any number of operations may have been in progress for the pod. Furthermore, some of those operations may have persisted only on the detached storage system, others may have persisted only on the storage system that remained in sync immediately after the detach was processed, and other operations may have persisted on both the detached storage system and the storage system that remained in sync. In this example, because the in-sync state for the pod does not record the operations that were persisted only on the detached storage system, any updates to the in-sync content and common metadata for the pod since the storage system detachment will not include those updates. This is why these updates would need to be undone—either explicitly by undoing the updates or implicitly by overwriting the content as part of the resynchronization process. For the synchronizing storage system itself, there may be two lists described before initiating reattachment of the detached storage system: (a) a list of operations that are in progress and remained in synchrony for any duration after the reattaching storage system was detached from the pod, sometimes referred to as the in-sync pending operations list, and (b) a list of changes to content or common metadata during the time period when the reattaching storage system was detached from the pod. Furthermore, depending on the implementation of the pod and storage system, the two lists associated with the synchronizing storage system may be represented by a single list, i.e., content that is not known to be on the reattaching storage system. In pods with multiple detached storage systems, and especially in pods where those storage systems are different, tracking changes since each detachment results in separate lists—and how those lists are described may vary significantly from one pod implementation to another. In some cases, an additional issue beyond tracking changes from the point of detachment and copying those changes to the attaching storage system is ensuring that new modification operations received during resynchronization are applied to the attaching storage system. Conceptually, this issue can be described as ensuring that the operations of copying data and processing modification operations received by the pod can be merged so that the result is properly up-to-date at the end of the attach and before the attaching storage system considers itself in sync for the pod.

[0421] With regard to single changed content resynchronization, one model for resynchronization is to generate a complete list of blocks that may differ between the synchronizing storage system and the attaching storage system—a list of detached blocks—and begin replicating any modification operations as they occur to the follower storage system. The complete list of blocks that may differ may include those from pending operations during synchronization at the time of detachment from the synchronizing storage system, pending operations at the time of detachment from the attaching storage system, and blocks known to have changed since detachment. The modification operations may store their modified content, as described, and resynchronization may proceed by locating ranges of blocks from the list of detached blocks and copying these blocks in sections from the synchronizing storage system to the attaching storage system. In this example, while copying a particular section, incoming modification operations that overlap the section being copied may be blocked during the copy, or provisions may be made to apply those modification operations after the section has been copied. This solution can create problems for virtual block range copy operations, such as virtualized implementations of the EXTENDED_COPY operation. Furthermore, the source range for the copy may not yet be resynchronized, and the target range may already be resynchronized. That is, a simple implementation of a virtual block range copy operation may not be able to copy the correct data to the target range (depending on the implementation) because the data is not known when the virtual block range copy operation is received, or the resynchronization operation itself may not be able to correctly resynchronize the target range because it may have assumed the target range was correctly synchronized when it was never resynchronized in its final form. However, there are several solutions to this problem. One solution is to not allow virtual block range copy operations during resynchronization. This can work in many cases because common uses of virtual block range copy operations—including client operations for file copy operations and virtual machine clone or later operations—typically respond to a virtual block range copy by copying the contents themselves directly through a sequence of read and write requests. Another solution is to remember the incomplete virtual range copy operation rather than performing a corrective operation that overwrites the target address range of any virtual address range copy operation and then performs the copy operation while accounting for the overwrite when the source data becomes available. Given that the resynchronization target may not know that the source data for the copy is incorrect, all such operations may have to be postponed until the full copy is complete. Optimizations can occur if the resynchronization target can be made aware of which areas have not yet been copied or when the resynchronization has finished processing a particular area of the volume.

[0422] Another aspect of resynchronizing a storage system may be updated block tracking. For example, keeping a list of all individual blocks modified while a storage system is detached (and then resynchronizing them individually) may not be practical in some cases because extended outage periods may result in a large number of blocks—some storage systems cannot read large collections of non-contiguous blocks very efficiently. Consequently, in some cases, it may be more practical to start a tracking range, such as a 1MB range of a volume, to reduce the amount of tracked metadata. This course-critical tracking may lag behind shorter-term activity tracking and may be preserved for as long as needed to handle resynchronization of any out-of-date storage system, regardless of whether it is down for minutes, hours, days, or weeks. With solid-state storage, as opposed to mechanical spinning storage, it may be quite practical to track which individual blocks of a volume, or a collection of volumes, or a complete pod, are modified, so that it is quite practical to resynchronize only those individual blocks that have changed. In general, there is little random read and write penalty and only a small penalty for reading from a multilevel map, resulting in a relatively easy merger of fine-grained activity as operations over short periods of time (e.g., in the range of 100 milliseconds to 10 seconds, or every few hundred operations or every few thousand operations) into a fine-grained map that names all modified blocks. Furthermore, the list of recent activity may be a list covering content modifications that have recently been recorded to a journal device (fast-write storage such as NVRAM with various flavors intended to support high write bandwidth and overwrite rates), although metadata about those modifications is likely kept in the journal for a longer period than the actual content. In this example, the merged list of all activity may be a bitmap where each bit represents a block or small group of blocks, or the merged list may be a list of block numbers or block ranges organized by the volume in a tree structure, such as a B-tree. Because nearby block numbers may be stored as the difference from one block number to another, such a list of block numbers may be easily compressed.

[0423] Resynchronizing storage systems may also include block tracking by tracking sequence numbers. For example, some storage systems associate a respective sequence number with each modification during normal operation—for every modification. In such cases, the last sequence number known to be synchronized with the detached storage system from the pod is all that is needed to query the in-sync storage system for the pod to find all content modified since detachment, including any content that may not have been replicated to the detached array around the time of detachment.

[0424] Resynchronizing the storage system may also include tracking changes as snapshots. For example, snapshots may be used to track changes since some time in the past, and the storage system may create a snapshot at the time of detachment by filtering out content that is not known to be complete. Alternatively, snapshots may be created regularly or with some periodicity, and the time of snapshot creation may be compared to the time of detachment to determine which snapshots may serve as a basis for resynchronizing the detached storage system. As a variation, snapshots created for the entire pod before detachment should exist on both the synchronizing storage system for the pod and the detached storage system and may be used in various ways for resynchronization. For example, the contents of the storage system being reattached may be reverted to its last synchronized snapshot prior to detachment and advanced from that point forward to match the currently synchronized content in the pod. Generally, snapshots show differences relative to a previous snapshot or show differences relative to the current content. Using these snapshot features, resynchronizing content to a reattaching storage system may involve replicating the difference between the time of reattachment and the time of the last fully synchronized snapshot before detachment. In some cases, resynchronization may use a snapshot-based model as a fallback. For example, short outage periods (such as outage periods on the order of a few minutes) may be handled through fine-grained tracking or record/playback operations that have occurred since the storage system detached, while longer outage periods may be handled by reverting to snapshots taken every few minutes—the threshold number of minutes may be a default value or may be specified by a user or administrator. Such a configuration may be practical because snapshots that occur more infrequently may have low long-term overhead but may generate more data to be restarted. For example—in some cases, a snapshot taken five minutes before detachment will transfer up to five minutes' worth of content modifications—a 10-second outage period may be handled by replaying the recorded operations, and resynchronization may occur in 10 seconds or less. In other cases, resynchronization after an outage period may be based on accumulated changes, for example, due to a limit on the accumulated size of the short-term map.

[0425] In some cases, resynchronization may be based on synchronous replication. For example, the snapshot-based resynchronization model described above may also support other resynchronization models. That is, a storage system that supports asynchronous or periodic replication may use a snapshot mechanism to replicate content during resynchronization. The asynchronous or periodic replication model may potentially copy outdated data during short outage periods, and the periodic replication model may be based on snapshot or checkpoint differential representations, which automatically handle outage periods. With asynchronous replication, there may be a reliance on snapshots or checkpoints as backups for extended outage periods, as described above. As a result, it may be practical to combine implementations or utilize available asynchronous or periodic replication implementations for resynchronization. However, one issue is that asynchronous or periodic replication models cannot be configured to keep the replication target up to date or fully in sync. As a result, with such resynchronization implementations, new in-flight operations may be applied, so that all modifications to the attached storage system can be tracked to be up to date for the pod.

[0426] In some cases, resynchronization may be implemented to include multi-stage resynchronization. For example, in a first stage, content up to a certain point may be replicated from the synchronizing storage system for the pod to the attaching storage system for the pod. In this example, a second snapshot may be taken during attach, and the difference between the first snapshot, which was the last snapshot known to be synchronized before detachment, and the difference between the first snapshot and the second snapshot is replicated to the attaching storage system. Such a mechanism may bring the attaching storage system more closely in sync than it was before attach, but it may still not be up to date. Thus, a third snapshot may be created, and the difference between the third snapshot and the second snapshot may be determined and then replicated to the attaching storage system. This third snapshot and the determined difference may constitute part of the difference between the content replicated up to the second snapshot and the current content. It is further contemplated that additional snapshots may be taken and replicated to get within seconds of being up to date. At this point, modification operations may be paused until the final snapshot is replicated—thereby bringing the attaching storage system up to date for the pod. In other cases, one or more snapshots may be switched, post-replication, to some mode in which modification operations received after the final resync snapshot are processed so that they can be merged with the replicated snapshot content. Such an implementation may include having the attaching storage system track those modification operations and applying them after snapshot replication is complete—or after snapshot replication is known to have synchronized the particular volume region affected by the particular modification operation. This implementation may have additional overhead, as tracking every operation until this underlying content is known to have been copied may result in a large number of tracked operations. An alternative is to consider content related to recently received operations, such as writes that depend on particular common metadata or extended copy operations from one block range to another, and requests that resync prioritize processing of that content or common metadata information. In this way, any received operations tied to content known to have been copied by such a process may then cause the tracking structure to be released much more quickly.

[0427] In some cases, resynchronization may be performed to use a directed acyclic graph of logical extents. As described above, replicated storage systems may be based on a directed acyclic graph of logical extents. In such storage systems, the resynchronization process may be expected to replicate the logical extent graph—including all leaf logical extent content—from the synchronizing storage system for a pod to the attaching storage system for the pod, ensuring that the graph is synchronized and remains synchronized before enabling the attaching storage system as a synchronizing pod member. Resynchronization in this model may proceed by having the target storage system for the attach retrieve top-level extent identities for each volume—for each file or object in a file- or object-based storage system. Any logical extent identities already known to the attach target may be considered current, while any unknown composite logical extents may be retrieved and then decomposed into elementary leaf logical extents or composite logical extents, each of which is either already known to the attach target or unknown to the attach target. Additionally, any unknown leaf logical extents may have their contents retrieved, or the identities of stored blocks may be retrieved to determine whether the blocks are already stored by the target storage system—then unrecognized blocks are retrieved from the synchronizing storage system. However, because some number of extents around the time of the storage system detachment may have the same identity but different content, such an approach may not necessarily lead to resynchronization, since only the operation that marked the logical extent read-only may form a new logical extent as a result of the modifying operation. Furthermore, ongoing modifying operations may have completed differently on different storage systems during the failure leading to the detachment, and if those modifying operations were on non-read-only logical extents, then those logical extents may have the same identity but different content at the two storage systems. However, some solutions include: when one collection of storage systems for a pod detaches another, the collection of storage systems may mark the leaf logical extents and composite logical extents associated with the ongoing modification operation and associate those leaf logical extents and composite logical extents with future reattach operations involving the detached storage system. Similarly, the reattaching storage system for the pod may identify the leaf logical extents and composite logical extents associated with the ongoing logical extent that it knew about. As a result, two sets of logical extents whose contents (for leaf extents) or references (for composite logical extents) need to be transferred in addition to transferring any unknown leaf logical extents or composite logical extents. Alternatively, coordinated snapshots may be taken periodically within the replicated pod, and the target of the reattach operation may ensure that logical extents created after the last coordinated snapshot are discarded or ignored during resynchronization. As yet another alternative, during the period in which a storage system is detached from a pod, the remaining synchronizing storage systems may produce snapshots that represent content with all completed operations and replay all potentially ongoing operations to apply to pod content occurring after the snapshot - with the result that any content not already replicated to the detached storage system is given a new logical extent identity that the detached storage system could never have received.

[0428] Another issue that may be faced in restart implementations is keeping extent graph-based synchronous replication fully synchronized and live and running. For example, resynchronization may proceed by first having the target storage system retrieve it from the in-sync storage system as described above, e.g., by transferring a more recent snapshot, such as the snapshot created at the beginning of the attach, and the target incrementally requests leaf logical extents and composite logical extents that it does not have. This process may include accounting for ongoing operations at the time of detachment, and at the end of this process, the contents up to that more recent snapshot are synchronized between the in-sync storage system for the pod and the attaching storage system. Further, this process may be repeated with another snapshot, and possibly additional snapshots, to bring the target storage system closer to the in-sync storage system. However, at some point, live data may also need to be transferred; to do this, live modification replication may be enabled for transfer to the attaching storage system after the last resynchronization snapshot, so that all modification operations not included in the snapshot may be delivered to the attaching storage system. This embodiment generates operations that describe modifications to leaf and composite logical extents included in the snapshot, and these descriptions may include creating new leaf and composite logical extents (with specified contents) or replacing existing leaf and composite logical extents with modified copies of those extents with new identities. If the description of the operation creates a new logical extent or replaces a logical extent already known to the attaching storage system, the operation may be processed normally as if the attaching storage system were in sync. If the description for the operation includes replacing at least one logical extent not already known to the attaching storage system, the operation may be made durable to allow completion, but full processing of the operation may be delayed until the replaced logical extent is received. Furthermore, to reduce the overhead associated with these operations waiting for such logical extent content transfer, the attaching storage system may prioritize those logical extents that are retrieved earlier than other logical extents. In this example, there may be no reason to transfer any series of snapshot images before allowing live operations, due to how the storage system can efficiently handle these operations waiting for such existing logical extents. Instead, a resync snapshot describing state information from the time of detachment (or from some point before detachment) could be transferred - operations would be handled as described above, with snapshots transferred from the synchronizing storage system to the attaching storage system, also as described above.

[0429] In some cases, a problem facing resynchronization implementations is preserving block references during resynchronization. For example, in a synchronously replicated storage system, a particular written block, or a particular set of blocks associated with an operation, may be given an identification that is included in the operation description for writing that block or set of blocks. In this example, a new write that replaces some or all of that block or set of blocks may provide a new identity for the block or set of blocks; this new identity may be constructed from a secure hash of the block contents (e.g., using SHA-256 or some other mechanism that has an appropriately very small probability of different blocks producing the same hash value), or the new identity may simply identify the write itself in a unique way, regardless of whether the two writes contained identical block contents. For example, the new identity may be a sequence number or a timestamp. Furthermore, if new identities for blocks or sets of blocks are shared in the distributed description of the write operation and stored in some map on each storage system as part of writing the block or set of blocks, then the logical extent may describe its contents in terms of these block or set of blocks identities. In such an embodiment, resynchronization of a leaf extent may reference blocks or sets of blocks already stored on the attaching storage system rather than transferring them from the synchronizing storage system. This embodiment may reduce the total data transferred during resynchronization. For example, data that was already written to the attaching storage system around the time of detachment but was not included in the resynchronization snapshot may have been stored with its identity, and the identity of that block or set of blocks may not need to be transferred again because it is already known and stored. Furthermore, if any number of virtual expansion copy operations caused the copying of block references between two leaf logical extents during the period when the storage system was detached, then the identity of the block or set of blocks may be used to ensure that the virtually copied blocks are not transferred twice.

[0430] In some cases, resynchronization implementations may use a content-addressed store in which stored blocks may have unique identities, which may be based on a secure hash of the block contents. In this example, resynchronization may proceed by transferring a list of all block identities related to the pod on the synchronizing storage system to the attaching storage system, along with a mapping of those block identities to volumes (or files or objects) within the pod. In this case, the attach operation—which may be integrated with the processing of live operations that change the volume-to-content mapping—may proceed by transferring from the synchronizing storage system for the pod those blocks that the attaching storage system does not know about. Furthermore, if any earlier version of the pod contents to block identities was known before the storage system detached from the pod, then the differences between that earlier version and the current version may be transferred instead of transferring the entire mapping.

Continuing with the example method shown in FIG. 21 , the example method includes identifying (2102) differences (2152) between an out-of-sync data set (2113) stored on an out-of-sync storage system (2124) and an in-sync data set (2112) stored on one or more in-sync storage systems (2114, 2128). Identifying (2102) differences (2152) between the out-of-sync data set (2113) stored on an out-of-sync storage system (2124) and an in-sync data set (2112) stored on one or more in-sync storage systems (2114, 2128) may be implemented using a variety of techniques, as described in detail above. One example technique includes one or more controllers using block tracking to keep a list of all individual blocks that are modified while any given storage system, such as storage system (2124), is detached, and as described in detail above. Other example techniques include using a combination of block tracking and snapshots, or block tracking by sequence number, among other techniques, as described in detail above.

[0432] In this example, block tracking is used to generate a list of all individual blocks modified at the in-sync storage system until the storage system (2124) is reattached. Block tracking begins when the storage system (2124) is detected as detached - this is when the next step, synchronizing (2104), described below, begins. Additionally, a "detached" storage system may be considered a storage system that is listed as a pod member but not listed as in-sync for the pod - a storage system that is listed as a pod member is in-sync if the pod member is online or currently available to consistently serve data for the pod. In this example, each storage system member of the pod may have its own copy of a membership list indicating the member storage systems for the pod, the membership list including which storage systems are currently known to be in-sync and which storage systems are included in the complete set of pod members. Generally, to be online for a pod, the membership list for a given storage system indicates that the given storage system is in sync for the pod, and the given storage system can communicate with all other storage systems in the membership list for the given storage system that are indicated as in sync for the pod. If a storage system cannot communicate with other storage systems indicated as in sync by the membership list, then the storage system stops processing incoming requests that modify datasets (or completes the requests with an error or error) until the storage system can verify that it is again in sync. A particular storage system may determine that a suspect storage system should be detached, allowing the particular storage system to continue operating based on being in sync with storage systems in its membership that are indicated as in sync. Furthermore, in this situation, to avoid a "split-brain" situation in which multiple isolated storage systems are processing I/O requests, the suspect storage system is prevented from continuing processing by the particular storage system, and the suspect storage system is prevented from requesting an intermediation service to determine which storage systems should continue processing I/O requests directed to the pod and which storage systems should stop processing I/O requests directed to the pod. Additional details for such an intermediation process are further described in Application Serial No. 15/703,559, which is incorporated by reference in its entirety.

21 also includes synchronizing (2104) the out-of-sync data set (2113) with the in-sync data set (2112) according to differences (2152) between the out-of-sync data set (2113) stored on the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128). Synchronizing (2104) the out-of-sync data set (2113) with the in-sync data set (2112) may be performed using multiple techniques, as described above. One example technique for resynchronization includes—given a list of modified blocks generated while the storage system (2124) was detached—replicating any modification operations corresponding to the list of modified blocks on the reattached storage system (2124). In this example, replicating the corrective operations may be performed similarly to how a follower storage system performs I/O operations provided by a leader storage system when the follower and leader storage systems are synchronized, as described above with respect to processing I/O operations. As described above, the resynchronization may also include corrective operations and may store correction comments, and the resynchronization may be performed by locating ranges of blocks from a list of detached blocks and copying those blocks to a section from the synchronizing storage system to the attaching storage system. Other example techniques for performing synchronization (2104) are described in more detail above.

[0434] The example method shown in FIG. 21 also includes re-establishing (2106) a synchronous replication relationship between an out-of-sync data set (2113) stored in an out-of-sync storage system (2124) and an in-sync data set (2112) stored in one or more in-sync storage systems (2114, 2128). Re-establishing (2106) the synchronous replication relationship between the out-of-sync dataset (2113) and the in-sync dataset (2112) may be performed similarly to initially establishing a synchronous replication relationship between storage systems when a pod is initially created, as described with respect to Figures 4 through 7 of Application Serial No. 15/713,153, which is incorporated by reference in its entirety, and includes identifying (2102) a plurality of storage systems across which the dataset is synchronously replicated; configuring (2104) one or more data communication links between each of the plurality of storage systems used to synchronously replicate the dataset; exchanging (2106) timing information for at least one of the plurality of storage systems among the plurality of storage systems; and establishing (2108) a synchronous replication lease in accordance with the timing information for at least one of the plurality of storage systems, wherein the synchronous replication lease identifies a period of time for which the synchronous replication relationship is valid.

[0435] For further explanation, FIG. 22 sets forth a flowchart illustrating an additional example method for resynchronization for storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. The example method shown in FIG. 22 also includes identifying (2102) differences (2152) between an out-of-sync dataset (2113) stored on an out-of-sync storage system (2124) and an in-sync dataset (2112) stored on one or more in-sync storage systems (2114, 2128), and resolving the differences (2152) between the out-of-sync dataset (2113) stored on the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128). The example method shown in FIG. 22 is similar to the example method shown in FIG. 21 because it includes synchronizing (2104) an out-of-sync dataset (2113) with an in-sync dataset (2112) in accordance with RFC 52, and re-establishing (2106) a synchronous replication relationship between the out-of-sync dataset (2113) stored in an out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

22 further includes identifying (2202) differences (2252) between metadata associated with the out-of-sync data set (2113) stored in the out-of-sync storage system (2124) and metadata associated with the in-sync data set (2112) stored in one or more in-sync storage systems (2114, 2128). Identifying (2202) differences (2252) between metadata associated with the out-of-sync data set (2113) stored in the out-of-sync storage system (2124) and metadata associated with the in-sync data set (2112) stored in one or more in-sync storage systems (2114, 2128) may be implemented using a variety of techniques, as described in more detail above. In addition to tracking modification operations that occur during the detached period of the resynchronizing storage system, as in the example described above, the storage system may also track metadata describing the modification operations tracked during the detached period of the resynchronizing storage system and store the tracked metadata in a log or journal device.

[0437] The example method shown in FIG. 22 further includes synchronizing (2204) the metadata associated with the out-of-sync dataset (2113) with the metadata associated with the in-sync dataset (2112) according to differences (2252) between the metadata associated with the out-of-sync dataset (2113) stored in the out-of-sync storage system (2124) and the metadata associated with the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128) - which may be implemented using a variety of techniques. One example technique includes one or more in-sync storage systems (2114, 2128) generating and maintaining a list of changes to metadata that have occurred since the storage system (2124) was detached, and the list of changes to metadata may be used to update the metadata representation of pod data at the out-of-sync storage system (2124) to bring the out-of-sync storage system (2124) in-sync, the in-sync state including a compatible graph to represent "common metadata," where metadata and common metadata are further described in Application Reference No. 15/66,418, which is incorporated by reference in its entirety. As described above, some operations persist in the detached storage system (2124) and result in modifications to the metadata associated with the out-of-sync dataset (2113) stored in the out-of-sync storage system (2124); these modifications to the metadata may be described in a dedicated list or in the same list of operations described above with respect to identifying (2102) differences between the out-of-sync dataset (2113) stored in the out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

[0438] For additional explanation, FIG. 23 sets forth a flowchart illustrating an additional example method of resynchronization for storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. The example method shown in FIG. 23 also includes identifying (2102) differences (2152) between an out-of-sync dataset (2113) stored on an out-of-sync storage system (2124) and an in-sync dataset (2112) stored on one or more in-sync storage systems (2114, 2128), and resolving the differences (2152) between the out-of-sync dataset (2113) stored on the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128). The example method illustrated in FIG. 23 is similar to the example method illustrated in FIG. 21 because it also includes synchronizing (2104) the out-of-sync dataset (2113) with the in-sync dataset (2112) in accordance with RFC 52, and re-establishing (2106) a synchronous replication relationship between the out-of-sync dataset (2113) stored in the out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

23 specifies, however, that identifying (2102) differences between an out-of-sync data set (2113) stored in an out-of-sync storage system (2124) and an in-sync data set (2112) stored in one or more in-sync storage systems (2114, 2128) further includes identifying (2302) one or more blocks (2352) in the out-of-sync data set (2113) that contain content that differs from the content stored in the in-sync data set (2112), which may be implemented using a variety of techniques. One example technique includes generating a list of blocks (2352) that may differ between the in-sync storage systems (2114, 2128) and the attached storage system (2124), as described in more detail above—one of the in-sync storage systems may initiate generation of the list of blocks in response to detaching or detecting that a particular storage system has detached. Further, detaching a storage system is described in more detail in Application Serial No. 15/696,418, which is incorporated herein by reference in its entirety.

[0440] The example method shown in FIG. 23 further specifies that synchronizing (2104) the out-of-sync dataset (2113) with the in-sync dataset according to differences (2152) between the out-of-sync dataset (2113) stored in the out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128) includes modifying (2304) one or more blocks in the out-of-sync dataset (2113) to match one or more blocks (2352) in the in-sync dataset (2113). Modifying (2304) one or more blocks in the out-of-sync data set (2113) to match one or more blocks (2352) in the in-sync data set (2113) may be performed—as described in more detail above—by locating ranges of memory addresses in one or more blocks (2352) from the list of detached blocks and copying content from those ranges of memory addresses from one of the in-sync storage systems (2114, 2128) to the attaching storage system (2124).

[0441] For additional explanation, FIG. 24 sets forth a flowchart illustrating an additional example method for resynchronization for storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. The example method shown in FIG. 24 also includes identifying (2102) differences (2152) between an out-of-sync dataset (2113) stored on an out-of-sync storage system (2124) and an in-sync dataset (2112) stored on one or more in-sync storage systems (2114, 2128), and resolving the differences (2152) between the out-of-sync dataset (2113) stored on the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128). The example method shown in FIG. 24 is similar to the example method shown in FIG. 21 because it also includes synchronizing (2104) the out-of-sync dataset (2113) with the in-sync dataset (2112) in accordance with RFC 52, and re-establishing (2106) a synchronous replication relationship between the out-of-sync dataset (2113) stored in the out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

[0442] However, the example method shown in FIG. 24 specifies that identifying (2102) differences between an out-of-sync data set (2113) stored in an out-of-sync storage system (2124) and an in-sync data set (2112) stored in one or more in-sync storage systems (2114, 2128) further includes identifying (2402), by at least one of the one or more in-sync storage systems (2114, 2128), one or more modifications (2452) to the data set (2112) that occurred since the out-of-sync storage system was detached. Identifying (2402), by at least one of the one or more in-sync storage systems (2114, 2128), one or more modifications (2452) to the data set (2112) that occurred since the out-of-sync storage system was detached may be performed by several techniques. As one example technique, one or more of the synchronizing storage systems (2124, 2128) may implement one of the block tracking techniques described above. Other example techniques include using a combination of block tracking and snapshots, or block tracking by sequence number, among other techniques, as described in detail above.

[0443] For further explanation, FIG. 25 illustrates a flowchart showing an additional example method of resynchronization for storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. The example shown in FIG. 25 also includes steps of identifying (2102) differences (2152) between an out-of-sync dataset (2113) stored on an out-of-sync storage system (2124) and an in-sync dataset (2112) stored on one or more in-sync storage systems (2114, 2128), and resolving the differences (2152) between the out-of-sync dataset (2113) stored on the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128). The example method shown in FIG. 25 is similar to the example method shown in FIG. 21 because it includes synchronizing (2104) an out-of-sync dataset (2113) with an in-sync dataset (2112) according to the method described above, and re-establishing (2106) a synchronous replication relationship between the out-of-sync dataset (2113) stored in an out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

[0444] However, the example method shown in FIG. 25 specifies that identifying (2102) differences between an out-of-sync data set (2113) stored in an out-of-sync storage system (2124) and an in-sync data set (2112) stored in one or more in-sync storage systems (2114, 2128) further includes identifying (2502), by at least one of the one or more in-sync storage systems (2114, 2128), one or more modifications (2552) to the data set (2112) that were outstanding when the out-of-sync storage system (2124) was detached. Identifying (2502), by at least one of the one or more in-sync storage systems (2114, 2128), one or more modifications (2552) to the data set (2112) that were outstanding when the out-of-sync storage system (2124) was detached may be performed using several techniques, as described in more detail above.

[0445] For further explanation, FIG. 26 illustrates a flowchart showing an additional example method for resynchronization for storage that synchronously replicates datasets in accordance with some embodiments of the present disclosure. The example method shown in FIG. 26 also includes identifying (2102) differences (2152) between an out-of-sync dataset (2113) stored on an out-of-sync storage system (2124) and an in-sync dataset (2112) stored on one or more in-sync storage systems (2114, 2128), and resolving the differences (2152) between the out-of-sync dataset (2113) stored on the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128). The example method shown in FIG. 26 is similar to the example method shown in FIG. 21 because it includes synchronizing (2104) an out-of-sync dataset (2113) with an in-sync dataset (2112) in accordance with RFC 52, and re-establishing (2106) a synchronous replication relationship between the out-of-sync dataset (2113) stored in an out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

[0446] However, the example method shown in FIG. 26 further includes detecting (2602) that an out-of-sync storage system (2124) has detached from the synchronous replication relationship between an out-of-sync dataset (2113) stored on the out-of-sync storage system (2124) and an in-sync dataset (2112) stored on one or more in-sync storage systems (2114, 2128), which may be implemented using a variety of techniques, as described in more detail above. The example method shown in FIG. 26 also includes tracking (2604) modifications (2652) to the dataset (2112) that have occurred since the out-of-sync storage system (2124) was detached, as described in more detail above.

[0447] For additional explanation, FIG. 27 sets forth a flowchart illustrating an additional example method for resynchronization for storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. The example method shown in FIG. 23 also includes identifying (2102) differences (2152) between an out-of-sync dataset (2113) stored on an out-of-sync storage system (2124) and an in-sync dataset (2112) stored on one or more in-sync storage systems (2114, 2128), and resolving the differences (2152) between the out-of-sync dataset (2113) stored on the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128). The example method shown in FIG. 23 is similar to the example method shown in FIG. 21 because it includes synchronizing (2104) an out-of-sync dataset (2113) with an in-sync dataset (2112) according to 2), and re-establishing (2106) a synchronous replication relationship between the out-of-sync dataset (2113) stored in an out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

[0448] However, the example method shown in FIG. 27 specifies that synchronizing (2104) the unsynchronized dataset (2113) with the in-synchronization dataset (2112) according to differences (2152) between the unsynchronized dataset (2113) stored in the unsynchronized storage system (2124) and one or more in-synchronization storage systems (2114, 2128) further includes replicating (2702) one or more snapshots of the one or more in-synchronization systems (2114, 2128) to the unsynchronized storage system (2124), which may be implemented using a variety of techniques, as described in more detail above.

2102) and the difference (2152) between the out-of-sync data set (2113) stored on the out-of-sync storage system (2124) and the in-sync data set (2112) stored on one or more in-sync storage systems (2114, 2128). The example method shown in FIG. 28 is similar to the example method shown in FIG. 21 because it includes synchronizing (2104) an out-of-sync dataset (2113) with an in-sync dataset (2112) in accordance with RFC 52, and re-establishing (2106) a synchronous replication relationship between the out-of-sync dataset (2113) stored in an out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

[0450] However, the example method shown in FIG. 28 specifies that synchronizing (2104) the out-of-sync dataset (2113) with the in-sync dataset (2112) according to differences (2152) between the out-of-sync dataset (2113) stored in the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128) further includes identifying (2802) one or more modifications (2852) of the dataset (2113) that were persisted only in the out-of-sync storage system (2124), which may be implemented using a variety of techniques, as described in more detail above, and undoing (2804) one or more modifications (2852) of the dataset (2113) that were persisted only in the out-of-sync storage system (2124).

29 illustrates a flowchart showing an additional example method of resynchronization for storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. The example method shown in FIG. 29 also includes identifying (2102) differences (2152) between an out-of-sync dataset (2113) stored on an out-of-sync storage system (2124) and an in-sync dataset (2112) stored on one or more in-sync storage systems (2114, 2128), and resolving the differences (2152) between the out-of-sync dataset (2113) stored on the out-of-sync storage system (2124) and one or more in-sync storage systems (2114, 2128). The example method shown in FIG. 29 is similar to the example method shown in FIG. 21 because it includes synchronizing (2104) an out-of-sync dataset (2113) with an in-sync dataset (2112) in accordance with RFC 52, and re-establishing (2106) a synchronous replication relationship between the out-of-sync dataset (2113) stored in an out-of-sync storage system (2124) and the in-sync dataset (2112) stored in one or more in-sync storage systems (2114, 2128).

[0452] However, the example method shown in FIG. 29 specifies that re-establishing (2106) the synchronous replication relationship between the out-of-sync data set (2113) stored in the out-of-sync storage system (2124) and the in-sync data set (2112) stored in one or more in-sync storage systems (2114, 2128) further includes enabling (2902) I/O operations for the database (2113) for the out-of-sync storage system (2124), which may be implemented using a variety of techniques, as described in more detail above.

[0453] For further explanation, FIG. 30 sets forth a flowchart illustrating an example method for managing connectivity to synchronously replicated storage systems (3014, 3024, 3028) in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (3014, 3024, 3028) illustrated in FIG. 30 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 30 may include the same components as the storage systems described above, fewer components, or additional components.

[0454] The example method shown in FIG. 30 includes identifying (3002) multiple storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated. The dataset (3012) shown in FIG. 30 may be embodied, for example, as the contents of a particular volume, as the contents of a particular shard of a volume, or as any other collection of one or more data elements. The dataset (3012) may be synchronized across the multiple storage systems (3014, 3024, 3028) such that each storage system (3014, 3024, 3028) maintains a local copy of the dataset (3012). In the examples described herein, such dataset (3012) is synchronously replicated across storage systems (3014, 3024, 3028) such that the dataset (3012) can be accessed through any of the storage systems (3014, 3024, 3028) that have performance characteristics such that any one storage system in the cluster will operate substantially less optimally than any other storage system in the cluster, at least as long as the cluster and the particular storage system being accessed are nominally operating. In such a system, modifications to the dataset (3012) should be made to copies of the dataset resident on each storage system (3014, 3024, 3028) so that accessing the dataset (3012) on any of the storage systems (3014, 3024, 3028) produces consistent results. For example, a write request issued to a dataset must be executed on all storage systems (3014, 3024, 3028) or must not be executed on any of the storage systems (3014, 3024, 3028). Similarly, some groups of operations (e.g., two write operations directed to the same location in a dataset) must be executed on all storage systems (3014, 3024, 3028) in the same order, or as if they were executed in the same order, so that the copies of the dataset resident on each storage system (3014, 3024, 3028) are ultimately identical on all storage systems (3014, 3024, 3028). Modifications to the dataset (3012) need not occur at exactly the same time, although some actions (e.g., issuing an acknowledgment that a write request has been directed to the dataset and that allows read access to locations within the dataset targeted by write requests that have not yet completed on all storage systems) may be deferred until the copies of the dataset (3012) on each storage system (3014, 3024, 3028) have been modified.

[0455] In contrast to processing write requests (or other requests that modify a dataset that is synchronously replicated across multiple storage systems), other types of requests may be serviced locally by the storage system that receives the request without increasing the distributed messaging latency of such operations. For example, read requests, query requests, or other requests that do not result in modification of the dataset (3012) can typically be processed locally by the storage system that receives the request without increasing the distributed messaging latency of such operations. For example, if a host issues a read request to a first storage system (3014) in a cluster of storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously distributed, then embodiments can ensure that there is no inline messaging between the storage systems (3014, 3024, 3028) that is required to complete the read request, which would frequently result in read latency that is identical to the latency of a local, non-replicated storage system. In some examples, such operations (e.g., read requests) may be blocked in an embodiment by configuring write requests (i.e., requests to write data to portions of a dataset that overlap with portions of the dataset read in response to a read request) or other forms of conflicting modification operations that have not yet completed at all of the storage systems (3014, 3024, 3028). Blocking may be necessary, for example, to preserve ordering requirements for multiple read requests that overlap in time with one or more concurrent modification requests. Such blocking can be used to ensure that a first read at one storage system, followed by a second read at another storage system in the pod, that coincides with a write operation or other modification operation on the same or another storage system in the pod that overlaps with the same write or other modification operation, returns content from before the modification operation for the second read, while never returning the results of the modification operation for the first read. This reverse chronological order for read operations can ensure that blocking of overlapping read requests due to in-flight modification operations that the storage system has learned about and that have not yet been processed anywhere in the pod does not occur by postponing a read request that may return results from an overlapping modification operation until it has determined that all other overlapping read requests have also returned results from their overlapping modification operations.

[0456] In the example method shown in FIG. 30, for a dataset (3012), identifying (3002) multiple storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated may be performed, for example, by examining a pod definition or similar data structure that associates the dataset (3012) with one or more storage systems (3014, 3024, 3028) that nominally store the dataset (3012). A 'pod,' as the term is used here and throughout the remainder of this application, may be implemented as a management entity representing a dataset, a collection of management objects and management operations, a collection of access operations to modify or read the dataset, and multiple storage systems. Such management operations may modify or query managed objects equally well across any of the storage systems, and access operations to read or modify the dataset may equally well operate across any of the storage systems. Each storage system may store a separate copy of the dataset as an appropriate subset of the dataset stored and advertised for use by the storage system, and operations that modify a management object or dataset performed and completed through any one storage system are reflected in subsequent management objects to query the pod or subsequent access operations to read the dataset. Additional details regarding "pods" may be found in previously filed provisional patent application Ser. No. 62/518,071, incorporated herein by reference. A storage system may be added to a pod, resulting in the pod's dataset (3012) being copied to that storage system and then kept up to date as the dataset (3012) is modified. A storage system may also be removed from a pod, resulting in the dataset (3012) no longer being kept up to date on the removed storage system. In such examples, the pod definition or similar data structure may be updated as storage systems are added to and removed from particular pods.

[0457] The example method shown in FIG. 30 also includes identifying (3004) a host (3032) that can issue I/O operations directed to the dataset (3012). The host (3032) shown in FIG. 30 may be implemented, for example, as an application server running external to the storage systems (3014, 3024, 3028) or as any other device that issues access requests (e.g., read, write) to the storage systems (3014, 3024, 3028) via one or more data communication paths. Identifying (3004) particular hosts (3032) that can issue I/O operations directed to the dataset (3012) may be performed, for example, by one or more of the storage systems (3014, 3024, 3028) maintaining a list or other data structure that includes the identification of each host from which the storage system (3014, 3024, 3028) has received an I/O operation directed to the dataset (3012), by consulting a list or other data structure that identifies each host that has the appropriate permissions necessary to access the dataset (3012), or in some other manner.

[0458] The example method shown in FIG. 30 also includes identifying (3006) multiple data communication paths (3022, 3026, 3030) between the host (3032) and multiple storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated. Each data communication path (3022, 3026, 3030) between the host (3032) and the multiple storage systems (3014, 3024, 3028) may represent a relationship between a host initiator port and a storage system target port, between a network interface on the host and a network interface on the storage system, etc. In such an example, there may be several host initiator ports and several storage system target ports, and the storage system may include several storage controllers, each of which may host multiple target ports. Target ports or network interfaces on separate storage systems should typically be distinct from each other, even within the same pod. Target ports may be managed using target port groups, which are groups of ports associated with storage system volumes that share common states with respect to active/optimized, active/non-optimized, standby, and offline. Target port groups may be associated with each storage controller of an individual storage system, rather than with the storage system as a whole. In fact, target port groups may be entirely arbitrary, including being associated with a subset of target ports even within a single storage controller. The storage system could also use host-initiated information when building or advertising target port groups. However, the storage system must provide this information consistently to each host-initiated program to ensure there is no confusion in the multipathing driver stack. In the example method shown in FIG. 30, identifying (3006) multiple data communication paths (3022, 3026, 3030) between a host (3032) and multiple storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated may be performed, for example, by use of the SCSI Asymmetric Logical Unit Access ("ALUA") mechanism described in more detail in the following paragraphs, by use of some other network discovery tool, or in some other manner.

30 also includes identifying (3008) one or more optimal paths from among a plurality of data communication paths (3022, 3026, 3030) between the host (3032) and a plurality of storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated. The storage system illustrated in FIG. 30 may identify (3008) one or more optimal paths from among a plurality of data communication paths (3022, 3026, 3030) between the host (3032), a plurality of storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated, and storage communication endpoints associated with the storage system. In the example method illustrated in FIG. 30, identifying (3008) one or more optimal paths may include identifying a single optimal path or identifying multiple optimal paths. For example, any path that meets various performance thresholds may be identified, a predetermined number of the most suitable paths (e.g., those paths that exhibit superior performance relative to other available paths) may be identified, a predetermined percentage of the most suitable paths may be identified, or a subset of more suitable paths (e.g., paths between a host and a particular storage system) may be identified. The reader will understand that because storage system hosts (3014, 3024, 3028) may be located at some distance from each other, because storage systems (3014, 3024, 3028) may be located on separate storage networks or separate portions of storage networks, or for some other reason, there may be a performance advantage associated with a host (3032) issuing I/O operations to one storage system versus another storage system. For example, there may be performance advantages associated with a host (3032) issuing I/O operations to a storage system physically located in the same data center or campus as the host (3032), compared to a host (3032) issuing I/O operations to a storage system physically located in a distant data center or campus. For reliability reasons, it may be beneficial to have the host (3032) maintain connectivity to all storage systems (3032), but for performance reasons, it may be preferable for the host (3032) to access a dataset (3012) through a specific storage system. The reader will understand that because different hosts may access the dataset (3012), one or more optimal paths for one host to access the dataset (3012) may differ from one or more optimal paths for another host to access the dataset (3012). In some embodiments, it is contemplated that two storage systems are sufficiently similar that paths to both storage systems may be deemed optimal. For example, if two storage systems are in the same data center or campus and have extensive networking between the host and those two storage systems, then the third storage system may be far enough away that it should not be used except as a fallback, but all paths between the host and two sufficiently similar storage systems may be candidates for being identified as an optimal path (3008).

[0460] In the example method shown in FIG. 30, identifying (3008) one or more optimal paths from among multiple data communication paths (3022, 3026, 3030) between a host (3032) and multiple storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated may be performed, for example, through the use of the SCSI ALUA mechanism. The SCSI ALUA mechanism is a SCSI specification that describes support for asynchronous access to storage system volumes (also known as "logical units" in SCSI) through multiple target ports as a set of commands and queries. This is described in the SPC-4 and SBC-3 technical standards. In such an embodiment, a volume (whose contents may represent a data set that is synchronously replicated across multiple storage systems) may report a unique ID to the host (3032) through multiple SCSI target ports associated with one or more target port groups, allowing the host (3032) to configure its I/O driver to access the volume through one or more SCSI host ports, and through some or all combinations of the host ports and advertised target ports. The unique volume ID is used by all volumes accessing the same volume. It may be used by the host (3032) I/O driver to recognize combinations of SCSI logical unit number, host port, and target port clauses. The host I/O driver can then issue SCSI commands down some, any, or all of these combinations (paths) to modify the state and contents of the identified volumes. Failures may result in the host reissuing requests down alternate paths, and performance concerns may lead the host (3032) to make heavy use of multiple paths to take advantage of improved host-to-storage system bandwidth through the use of multiple ports and multiple network interconnects.

[0461] The ALUA specification for SCSI can be used to describe multiple target port groups that can access a volume, each of which may be assigned a state. A target port group may represent one or more SCSI ports on a storage system. In a multi-controller storage system, a target port group could represent all SCSI target ports on one controller, or with a symmetrically accessible, synchronously replicated storage system, a target port group may represent all SCSI target ports on an individual storage system, or target ports may be grouped in some other way. The states that can be associated with a target port group indicate whether the port group should be preferred for issuing I/O (active/optimized), not preferred for issuing I/O (active/non-optimized), standby (I/O cannot be issued until there is a state change back to active/optimized or active/non-optimized), or whether the target port may be offline, such as for not responding. The SCSI specification allows the definition of target port groups and state-assigned ALUA target port groups to be specific not only to each volume, but also to each requesting host (or even to each requesting host port), allowing for each volume the storage system to present a unique set of target port groups and state-assigned target port groups to each host or host port that has access to that volume.

[0462] With symmetrically accessible, synchronously replicated storage systems, all storage systems in a pod can present the same volumes to the host as if all storage systems in the pod were a single storage system presenting the same volumes through SCSI target ports on some or all storage systems for the pod. These mechanisms can then provide all the desired functionality for directing and redirecting access to volumes within the pod. For a host that receives better performance for a particular storage system for the pod (e.g., due to network or geographic proximity of the host to that storage system), the ALUA target port group assigned a state for that storage system for that host's host ports may be indicated as active/optimized. Meanwhile, for other hosts that receive less performance for that particular storage system for the pod, the ALUA target port group assigned a state for that storage system for other other host ports may be indicated as active/non-optimized. In this manner, members of the target port group that are determined to be active/optimized may be identified to the host as the optimal path(s) (3008).

[0463] When a new storage system is added to a pod, then a new target port group may be added for each volume in the pod to the host ports that access that volume, and the target port group may be assigned the appropriate state for host/storage system proximity for the new storage system. After some number of SAN-level events, the host may recognize the new ports for each volume and configure its driver to use the new paths appropriately. The storage system, on behalf of the pod, may monitor for host access (e.g., wait for REPORT LUNS and INQUIRY commands) to determine that the host is now appropriately configured to use the SCSI target ports on the newly added storage system. If a storage system is removed from a pod, then other storage systems remaining in the pod may stop reporting any target ports or target port groups to any host ports for the removed storage system for the pod's volumes. Additionally, the removed storage system may stop listing the pod's volumes in any REPORT LUN requests, and the storage system may begin reporting that the volume is not present in response to commands for the pod's volumes. If a volume is moved into or out of a pod, resulting in the volume being associated with a larger or smaller collection of storage systems, the same actions can be applied to individual volumes that would be applied when adding or removing a storage system from a pod. With regard to failure handling, host I/O drivers access their volumes through target ports in the target port group that are assigned as active/optimized if any such paths are available and functioning properly, but can switch to an active/non-optimized path if the active/non-optimized path is unavailable or not functioning properly.

[0464] Also, in the example method shown in FIG. 30, identifying (3008) one or more optimal paths from among multiple data communication paths (3022, 3026, 3030) between a host (3032) and multiple storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated may be performed in an automated manner by using timing information or network information to determine, for example, that a host path to a particular interface or storage system within a pod has lower latency, better throughput, or less switching infrastructure than a host path to another particular interface or storage system within a pod. Thus, in such an example, identifying (3008) one or more optimal paths from among a plurality of data communication paths (3022, 3026, 3030) between a host (3032) and a plurality of storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated includes identifying one or more data communication paths that exhibit a relatively lowest latency between the host and one of the storage systems, identifying one or more data communication paths that exhibit queuing between the host and one of the storage systems that is below a predetermined threshold, and the like.

[0465] For example, in IP-based networks, ping and traceroute commands (or leveraging their basic ICMP echo requests) may be used to determine latency and network routes between known host network interfaces and storage systems within a pod. The traceroute facility, or direct use of ICMP echo requests with incrementally increasing TTL fields (to limit the number of network hops before the route sends an ICMP Time Exceeded response), can be used to determine the presence or absence of particular network hops with higher latency. In this case, host interface to storage interface routes with high latency hops may be avoided in favor of host interface to storage interface routes without high latency hops. Alternatively, if there is a network route with fewer hops and lower latency than others, the storage system with the network interface may be preferred.

[0466] In Fibre Channel-based networks, the HBA API specification of the plug-in library supported by the Storage Networking Industry Association from its Fibre Channel Working Group can be used to precisely plan FC storage networks. The ELS echo function of the Fibre Channel protocol can also be used to detect network latency. Similar to the IP network described above, this can be used to identify host ports to target port networks that have lower latency and fewer network hops than other host and target port combinations, which can be used to determine which storage systems within a pod are closer to or better connected to one host or another in order to configure the storage systems within the pod for active/optimized versus active/non-optimized per host.

[0467] The example method illustrated in FIG. 30 also includes indicating (3010) to the host (3032) the identification of one or more optimal paths. In the example method illustrated in FIG. 30, the storage systems (3014, 3024, 3028) may indicate (3010) the identification of the optimal paths to the host (3032), for example, via one or more messages exchanged between the storage systems (3014, 3024, 3028) and the host (3032). Such messages may be exchanged using many of the mechanisms described above and may identify the optimal paths through the use of a port identifier, a network interface identifier, or some other identifier. For example, indicating (3010) the identification of one or more optimal paths to the host (3032) may be implemented by indicating (434) one or more active/optimized paths to the host (3032) via a SCSI ALUA mechanism.

[0468] The reader will understand that the storage systems (3014, 3024, 3028) described herein may utilize host definitions to define hosts (3032) as named collections of ports or network interfaces, and that these host definitions may include additional information or characteristics, such as operating system or application type or workload classification. Host definitions may be named for management convenience and represented as first-class objects in the storage system user interface, or may be grouped together in various ways to list all hosts associated with a particular application, user, or host-based database or file system cluster. These host definitions may serve as convenient management objects for the storage system to associate information about host location or for host-to-storage system prioritization for pods. It may be convenient for the storage system to manage one host definition per host rather than one host definition per pod. This may be convenient because the initiator ports and network interfaces associated with a host are likely the same for all pods. This may not be the case if pods are used as a strong form of virtual device where each pod is securely isolated from other pods, but this may be convenient and easier to set up for any application or implementation that falls short of such secure pod isolation.

[0469] If a pod supplies datasets or storage objects to the same host from multiple storage systems within the pod, and if the ALUA state and host-to-storage system priorities for that host must be managed in a coordinated manner across all storage systems for the pod, then the host definition may need to be coordinated or synchronized across the pod; however, unlike most other managed objects for the pod, the host may be a storage system object (because network interfaces and SCSI target ports are often storage system objects) rather than a pod object. As a result, host objects may not be easily synchronized between pod members because the definitions may not be consistent.

[0470] Additionally, a host may be interconnected to one storage system for a pod through one set of host-side initiators and network interfaces, and to another storage system for a pod through a different set of host-side initiators and network interfaces. Furthermore, there may be some overlap between the two sets, or there may be no overlap between the two sets. In some cases, there may be host information that the storage system can use to determine that the interfaces represent the same host. For example, the interfaces may use the same iSCSI IQN, or the host-side driver may provide host information to the storage system to indicate that various initiators or network interfaces represent the same host. In other cases, such information may not be available. In the absence of discoverable information, host-defined parameters may instead be provided to the storage system by the user or through some API or other interface to associate a host name with a set of network endpoints, iSCSI IQNs, initiator ports, etc.

[0471] If a portion of the dataset associated with a pod is exported to a specific host through a host definition (i.e., it is provided to the host based on the host definition from the pod's current storage system's unique network endpoints and SCSI targets through a list of network endpoints, iSCSI IQNs, or initiator ports), then when an additional storage system is added to the pod, the host definition of the added storage system can be examined. If no host with the same host object name or a host with a duplicate list of host network endpoints, iSCSI IQNs, or initiator ports exists in the added storage system, then the host definition can be copied to the added storage system. If a host definition with the same names and the same configuration of host network endpoints, iSCSI IQNs, and initiator ports exists in the added storage system, then the host definition from the original pod member storage system and the host definition of the added storage system can subsequently be linked and reconciled. If hosts with the same name but different configurations of host network endpoints, iSCSI IQNs, or initiator ports exist in the added storage systems, then qualified versions of the host object can be exchanged between the storage systems for the pod, with the different versions named with the storage system qualifier. For example, storage system A may present its host definition as A:H1, while storage system B may present its host definition as B:H1. The same may also be provided for host definitions that have different names but some overlap in host network endpoints, iSCSI IQNs, or initiator ports. In that case, the host definitions may not be copied between storage systems, but instead remain local to the storage systems, resulting in, for example, host definition A:H1 listing host initiators X and Y, and host definition B:H2 listing host initiators Y and X. Furthermore, operations may be provided to synchronize these host definitions. If two host definitions have the same name or an overlapping set of host network endpoints, iSCSI IQNs, or initiator ports, then the user may be provided with a simple interface to unify them with a common name, the exchanged host network endpoints, iSCSI IQNs, or initiator ports then linked together. If the only conflict between two definitions is that some host definitions contain a host interface that is not listed in another, but at least one host interface matches, then these definitions could be merged and linked automatically, rather than waiting for the user to make such a request.

[0472] If a storage system in a pod detaches another storage system as a result of a set of conditions (failure, outage, etc.), then the detached storage system may make changes to its host definitions if it is offline for the pod but otherwise still running. Also, a storage system that remains online for the pod may make changes to its host definitions. The result may be inconsistent host definitions. If the detached storage system is later reattached to the pod, then the host definitions may no longer match. At that point, the pod may resume reporting host definitions with its storage system name prefix to distinguish between the separate definitions on each storage system.

[0473] Another aspect of the host definition is that it may define location or storage system priorities as part of configuring what is returned for ALUA information in terms of which storage system target ports should result in an active/optimized ALUA status and which should result in an active/non-optimized ALUA status. This state may also need to be reconciled and linked between storage systems. If so, that may be another aspect that requires reconciliation. Conflicts, or the lack of any settings, detected when reconciling host definitions for storage systems in a pod may also present an opportunity to prompt the user to set location or storage system priorities.

[0474] The reader should note that because a host definition may actually be a storage-system-level object rather than a pod-level object, the same host definition can be used not only for pods that do not span more than a single storage system, but also for pods that span multiple storage systems. The use of host in the context of a purely local pod (i.e., one pod that spans a different set of storage systems than another pod) could alter how the host (or list of hosts) is viewed. For example, in the context of a local pod, it may not make sense to qualify hosts by storage system, and it may not make sense to list hosts that do not have a path to the local storage system. This example could be extended to pods with different member storage systems. For example, a host definition may conflict with a paired storage system for one pod (making storage system qualification useful). On the other hand, a host definition may not conflict with a different paired storage system for a different pod (making storage system qualification potentially unnecessary). Similar issues can arise with host usage when one pod is spread across one set of additional storage systems and another pod is spread across another set of additional storage systems, where only the target interfaces for the storage systems relevant to the particular pod's issue, and only the host endpoints, iSCSI IQNs, and initiator ports are visible to the target interfaces on the relevant storage systems for the pod.

[0475] While only one of the storage systems (3014, 3024, 3028) shown in FIG. 30 is explicitly shown as performing the steps described above, the reader will understand that each of the storage systems (3014, 3024, 3028) may perform the steps described above. In practice, each of the storage systems (3014, 3024, 3028) may perform the steps described above substantially simultaneously, making the identification of the optimal path a coordinated effort. For example, each storage system (3014, 3024, 3028) may individually identify all data communication paths between itself and the host, collect various performance metrics associated with each data communication between itself and the host, and share such information with the other storage systems in an effort to identify one or more optimal paths.

[0476] For further explanation, FIG. 31 sets forth a flowchart illustrating an additional example method for managing connectivity to synchronously replicated storage systems (3014, 3024, 3028) in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (3014, 3024, 3028) illustrated in FIG. 31 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 31 may include the same components, fewer components, or additional components as the storage systems described above.

[0477] The example method shown in FIG. 31 includes identifying (3002) a plurality of storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated; identifying (3004) a host (3032) that can issue I/O operations directed to the dataset (3012); and identifying a plurality of data communication paths (3022, 3032) between the host (3032) and the plurality of storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated. The example method illustrated in FIG. 31 is similar to the example method illustrated in FIG. 30 in that it includes identifying (3006) one or more optimal paths from among a plurality of data communication paths (3022, 3026, 3030) between a host (3032) and a plurality of storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated, and indicating (3010) to the host (3032) the identification of the one or more optimal paths.

31 also includes detecting (3102) that the dataset (3012) is synchronously replicated across the updated set of storage systems. In the example method shown in FIG. 31, the set of storage systems across which the dataset (3012) is synchronously replicated may change for various reasons. For example, the set of storage systems across which the dataset (3012) is synchronously replicated may change because one or more properly functioning storage systems are added to or removed from a pod. Furthermore, the set of storage systems across which the dataset (3012) is synchronously replicated may change because one or more storage systems become unreachable and detached from a pod in response to becoming unreachable or unavailable. In the example shown in FIG. 31, detecting (3102) that the dataset (3012) is synchronously replicated across the updated set of storage systems may be performed, for example, by detecting a change to a pod definition, by detecting that a storage system has become unreachable or otherwise unavailable, or in some other manner.

31 illustrates an embodiment in which a data set (3012) is synchronously replicated across an updated collection of storage systems. In such an example, storage system (3024) was initially identified (3002) as one of multiple storage systems (3014, 3024, 3028) across which data set (3012) is synchronously replicated, as described in FIG. 30. However, the example illustrated in FIG. 31 illustrates an embodiment in which storage system (3024) becomes unreachable or otherwise unavailable, indicated herein by the use of dotted lines (representing no active connection) for data communication links (3016, 3018, 3020) between other storage systems (3014, 3028) and for data communication paths (3030) that can be used to couple storage system (3024) with a host (3032) for data communication with storage system (3024). Thus, the updated set of storage systems across which dataset (3012) is synchronously replicated includes only two storage systems (3014, 3028), whereas the initial set of storage systems across which dataset (3012) was synchronously replicated included all of the storage systems shown (3014, 3024, 3028).

[0480] The example method shown in FIG. 31 also includes identifying (3104) multiple data communication paths between the host (3032) and the updated set of storage systems (3014, 3028). In the example method shown in FIG. 31, identifying (3104) multiple data communication paths between the host (3032) and the updated set of storage systems (3014, 3028) may be performed, for example, by using the SCSI ALUA mechanism described in more detail above, by using some other network discovery tool, or in some other manner.

31 also includes identifying (3106) an updated set of optimal paths from among the multiple data communication paths between the host (3032) and the updated set of storage systems. In the example method illustrated in FIG. 31, identifying (3106) the updated set of optimal paths may include identifying a single optimal path or identifying multiple optimal paths. Any path that meets various performance thresholds may be identified, a predetermined number of the most optimal paths (e.g., those paths that exhibit superior performance relative to other available paths) may be identified, a predetermined percentage of the most optimal paths may be identified, or a subset of more optimal paths (e.g., paths between a host and a particular storage system) may be identified. The reader will understand that because storage system hosts (3014, 3024, 3028) may be located at some distance from one another, because the storage systems (3014, 3024, 3028) may be located on separate storage networks or separate portions of a storage network, or for some other reason, there may be a performance advantage associated with a host (3032) issuing I/O operations to one storage system versus another storage system. For example, there may be a performance advantage associated with a host (3032) issuing I/O operations to a storage system that is physically located in the same data center or campus as the host (3032), compared to a host (3032) issuing I/O operations to a storage system that is physically located in a distant data center or campus. For reliability reasons, it may be beneficial for the host (3032) to maintain connectivity to all storage systems (3032), but for performance reasons, it may be preferable for the host (3032) to access a dataset (3012) through a specific storage system. The reader will understand that because different hosts may access the dataset (3012), one or more optimal paths for one host to access the dataset (3012) may differ from one or more optimal paths for another host to access the dataset (3012). In the example method shown in FIG. 30, identifying (3106) an updated set of optimal paths from among multiple data communication paths between the host (3032) and the updated set of storage systems may be performed in an automated manner, by using timing or network information to determine that a host path to a storage system within a particular interface or pod has lower latency, better throughput, or less switching infrastructure than a host path to a storage system within another particular interface or pod, for example, by using the SCSI ALUA mechanism described above, or in some other manner. The reader will understand that while several of the above paragraphs refer to a "set," such a set may include a single member and does not impose any particular restrictions on how such a set may be represented.

[0482] The reader will understand that there can be various sources of the performance advantage described above. In the case of a write, a host requesting a write to a more distant storage system requires long-distance bandwidth for the network from the host to the distant storage system, in addition to the long-distance bandwidth required (in any case) for the storage-system-to-storage-system replication itself (which uses host-to-storage bandwidth that would not otherwise be needed, or which generates traffic for the write content in both directions when traffic would only be needed in one direction). Furthermore, in the case of a write, if long-distance latency is significant, then that latency is multiplied by four or six (the first or third phase of a two- or four-phase write from the host to the distant storage system, plus the delivery of the write content from the distant storage system to the local storage system, plus the delivery of a completion or similar indication from the local storage system to the distant storage system, plus the final completion sent from the distant storage system to the host for the last part of the two- or four-phase write request). In contrast, for writes to the local storage system, long-distance latency is incurred only twice: once for delivery of the write contents from the local storage system to the distant storage system, and once for delivery of a completion or similar indication from the distant storage system to the local storage system. For reads, a host requesting a read from the local storage system often does not consume any long-distance bandwidth and typically does not incur a long-distance latency penalty.

[0483] The example method shown in FIG. 31 also includes indicating (3108) to the host (3012) the identification of the updated optimal path. In the example method shown in FIG. 31, the storage systems (3014, 3024, 3028) may indicate (3018) the identification of the updated optimal path to the host (3032), for example, through one or more messages exchanged between the storage systems (3014, 3024, 3028) and the host (3032). Such messages may be exchanged using many of the mechanisms described above and may identify the optimal path through the use of a port identifier, a network interface identifier, or some other identifier. The reader will note that in some embodiments, part of the process by which a storage system (3014, 3024, 3028) may indicate (3108) an updated identification of an optimal path to a host (3032) may include piggybacking such information onto a response to a command issued by the host (3012). For example, one of the storage systems (3014, 3024, 3028) may gather SCSI unit attention for the host (3110). SCSI unit attention is a mechanism that allows a device (e.g., a storage system) to inform a host-side SCSI driver that the device's operational state or fabric state has changed. In other words, by gathering unit attention, the storage system may indicate to the host that the host should query the storage system for a state change, which allows the host to notice that the target port group state has changed to indicate different sets of active/optimized and active/non-optimized target port groups. In such an example, the target (e.g., storage system) internally collects a "unit attention" that is returned to the host (3012) when a response to a command instructing the host-side SCSI driver to request an updated ALUA state before clearing the unit attention is sent to the host (3012). This mechanism may allow the storage system to have the host update its ALUA state as desired, but it is dependent on a future SCSI request of some kind being issued to any target port that is not offline. In such an example, because the SCSI protocol takes the form of commands issued by the host (3012) and responses returned by the target (e.g., storage system), the SCSI "unit attention" mechanism provides a way for the target to return unsolicited updates to the host (3012), so transferring information to update the optimal path may need to be piggybacked on this mechanism in a slightly circuitous manner.

[0484] For further explanation, FIG. 32 sets forth a flowchart illustrating an additional example method for managing connectivity to synchronously replicated storage systems (3014, 3024, 3028) in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (3014, 3024, 3028, 3210) illustrated in FIG. 32 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 32 may include the same components, fewer components, or additional components as the storage systems described above.

[0485] The example method shown in FIG. 32 includes identifying (3002) a plurality of storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated; identifying (3004) a host (3032) that can issue I/O operations directed to the dataset (3012); identifying (3006) a plurality of data communication paths (3022, 3026, 3030) between the host (3032) and the plurality of storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated; The example method illustrated in FIG. 32 is similar to the example methods illustrated in FIGS. 30 and 31, in that it also includes identifying one or more optimal paths from the host (3032) (3008), and indicating to the host (3032) the identification of the one or more optimal paths (3010), detecting that the dataset (3012) is synchronously replicated across the updated set of storage systems (3102), identifying multiple data communication paths between the host (3032) and the updated set of storage systems (3014, 3028) (3104), identifying an updated set of optimal paths from among the multiple data communication paths between the host (3032) and the updated set of storage systems (3106), and indicating to the host (3012) the identification of the updated optimal paths (3108).

32, detecting (3102) that dataset (3012) is synchronously replicated across the updated set of storage systems may include detecting (3202) that storage system (3024) has detached from the original set of storage systems across which dataset (3012) was synchronously replicated. In the example method shown in FIG. 32, storage system (3024) may be considered 'detached' when storage system (3024) is no longer involved in synchronous replication of dataset (3012) across multiple storage systems. A particular storage system may detach for example, due to a hardware failure within the storage system, due to a network failure that prevents the storage system from engaging in data communications, due to a loss of power to the storage system, due to a software conflict on the storage system, or for a variety of other reasons. In the example method shown in FIG. 32, detecting (3202) that storage system (3024) has detached from the original set of storage systems across which data set (3012) was synchronously replicated may be performed, for example, by determining that the storage system has become unavailable or otherwise unreachable. In the example method shown in FIG. 32, one of the storage systems (3024) is shown as detached due to a networking failure that prevents the storage system from engaging in data communications because all data communication links (3016, 3018, 3020, 3208) and data communication paths (3030) used by storage system (3024) are represented by dotted lines to indicate that the data communication links (3016, 3018, 3020, 3208) and data communication paths (3030) used by storage system (3024) are not operational.

32, detecting (3102) that dataset (3012) is synchronously replicated across the updated set of storage systems may also include detecting (3204) that a storage system (3210) that was not included in the original set of storage systems across which dataset (3012) was synchronously replicated has attached to the set of storage systems across which dataset (3012) is synchronously replicated. In the example method shown in FIG. 32, a storage system (3210) may be considered "attached" when it participates in the synchronous replication of dataset (3012) across multiple storage systems. A particular storage system may be attached, for example, because the storage system is added to a pod, because the storage system recovers from a hardware failure within the storage system, because the storage system recovers from a network failure, because the storage system recovers from a loss of power to the storage system, because the storage system recovers from a software collision on the storage system, or for various other reasons. In the example method shown in FIG. 32, one of the storage systems (3210) not included in any of the previous figures is shown as attached to a collection of storage systems across which data set (3012) is synchronously replicated, and storage system (3210) is coupled for data communication with host (3032) and other storage systems (3024) via one or more data communication links (3206) and data communication paths (3212). The reader will understand that while data communication links are not shown between some of the storage systems (3028, 3210), such data communication links may actually exist but are merely omitted here for convenience of explanation.

[0488] The example method shown in FIG. 32 may also include monitoring (3206) host access to a storage system (3210) that was not included in the original set of storage systems across which the dataset (3012) was synchronously replicated. As described above, if a new storage system is added to a pod, then a new target port group may be added for each volume in the pod to the host ports that access the volume, and the target port group may be assigned a state appropriate to the host/storage system proximity for the new storage system. After some number of SAN-level events, the host may recognize the new ports for each volume and configure its driver to appropriately use the new paths. The storage system, on behalf of the pod, may monitor for host access (e.g., by waiting for REPORT LUNS and INQUIRY commands) to determine that the host is now appropriately configured to use the SCSI target ports on the newly added storage system. In such an example, a host may be monitored for assurance that the host is ready to issue commands to a newly added member of a target port group before taking action against a pod that depends on the host being ready to issue commands to members of that target port group. This may be useful, for example, when coordinating the removal of a pod member. In such an example, if one or more hosts known to be using the storage system being removed are not yet aware that they are using a recently added storage system, and the member being removed is the last remaining storage system known to serve one or more of those hosts, then it may be beneficial to issue an alert before allowing the operation to proceed (or may immediately prevent the operation).

[0489] For further explanation, FIG. 33 sets forth a flowchart illustrating an additional example method for managing connectivity to synchronously replicated storage systems (3014, 3024, 3028) according to some embodiments of the present disclosure. While not shown in great detail, the storage systems (3014, 3024, 3028, 3210) shown in FIG. 33 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system shown in FIG. 33 may include the same components as the storage systems described above, fewer components, or additional components.

[0490] The example method shown in FIG. 33 includes identifying (3002) a plurality of storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated; identifying (3004) a host (3032) that can issue I/O operations directed to the dataset (3012); and identifying a plurality of data communication paths (3022, 3024) between the host (3032) and the plurality of storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated. 30, 31, and 32, as it may include identifying (3006) one or more optimal paths from among a plurality of data communication paths (3022, 3026, 3030) between a host (3032) and a plurality of storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated; identifying (3008) one or more optimal paths from among a plurality of data communication paths (3022, 3026, 3030) between the host (3032) and a plurality of storage systems (3014, 3024, 3028) across which a dataset (3012) is synchronously replicated; and indicating (3010) to the host (3032) the identification of the one or more optimal paths.

33 also includes detecting (3302) changes to at least one or more data communication paths (3022, 3026, 3030) between the host (3032) and the plurality of storage systems (3014, 3024, 3028) to which the dataset (3012) is synchronously replicated. Detecting (3302) changes to at least one or more data communication paths (3022, 3026, 3030) between the host (3032) and the plurality of storage systems (3014, 3024, 3028) to which the dataset (3012) is synchronously replicated may be performed, for example, by detecting that a particular data communication path is no longer operational, by determining that performance (e.g., bandwidth, throughput) across a particular data communication path has changed by more than a predetermined threshold amount, by determining that fewer or additional hops have been introduced to a particular data communication path, etc. The reader will understand that changes to the multiple data communication paths (3022, 3026, 3030) between the host (3032) and the multiple storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated may affect which particular data communication path is identified as the optimal path, and therefore the storage systems (3014, 3024, 3028) may need to repeat the step (3008) of identifying an optimal path from among the multiple data communication paths (3022, 3026, 3030) between the host (3032) and the multiple storage systems (3014, 3024, 3028) across which the dataset (3012) is synchronously replicated, and the step (3010) of indicating the identification of one or more optimal paths to the host (3032).

[0492] The example method shown in FIG. 33 also includes detecting (3304) a change to host (3032). The change to host (3032) may occur, for example, as a result of a software or hardware upgrade to host (3032), as a result of a loss of power to host (3032), as a result of a hardware or software failure at host (3032), as a result of host (3032) being moved, as a result of a new host being used to support the execution of some application that issues I/O operations directed to dataset (3012), or for a variety of other reasons. The reader will understand that changes to the host (3032) may affect which particular data communication path is identified as the optimal path, and therefore the storage systems (3014, 3024, 3028) may need to repeat the step (3008) of identifying an optimal path from among multiple data communication paths (3022, 3026, 3030) between the host (3032) and multiple storage systems (3014, 3024, 3028) across which the data set (3012) is synchronously replicated, and the step (3010) of indicating to the host (3032) the identification of one or more optimal paths.

[0493] For further explanation, FIG. 34 sets forth a flowchart illustrating an additional example method for managing connectivity to synchronously replicated data across storage systems (3424, 3426, 3428) in accordance with an embodiment of the present disclosure. While not shown in great detail, the storage systems (3424, 3426, 3428) illustrated in FIG. 34 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 34 may include the same components as the storage systems described above, fewer components, or additional components. Furthermore, each of the storage systems (3424, 3426, 3428) illustrated in FIG. 34 may be connected to each other via one or more data communication links (3420, 3422) and to the host (3402) via one or more data communication paths (3410, 3412, 3414).

34 includes receiving (3404) an I/O operation (3416) directed to a dataset (3418) that is synchronously replicated across multiple storage systems (3424, 3426, 3428). In the example method shown in FIG. 34, a host (3402) receives (3404) an I/O operation (3416) directed to the dataset (3418) that is synchronously replicated across multiple storage systems (3424, 3426, 3428), for example, from an application running on the host, as a result of some user interaction with the host (3402), or in various other ways. An I/O operation (3416) directed to a dataset (3418) that is synchronously replicated across multiple storage systems (3424, 3426, 3428) may be implemented, for example, as a request to write data to the dataset (3418), as a request to read data from the dataset (3418), as a request to copy the data in the dataset (3418) and store such a copy elsewhere, as a request to take a snapshot of the data in the dataset (3418), etc.

34 also includes identifying (3406) a particular storage system (3426) among the plurality of storage systems (3424, 3426, 3428) as a preferred storage system for receiving the I/O operation (3416). In the example method shown in FIG. 34, the host (3402) may identify (3406) a particular storage system (3426) among the plurality of storage systems (3424, 3426, 3428) as a preferred storage system for receiving the I/O operation (3416), for example, by tracking (or otherwise having access to information describing the response times) response times previously experienced by the host (3402) when issuing I/O operations to each of the storage systems (3242, 3426, 3428) and selecting the storage system (3426) that exhibited the fastest response time as the preferred storage system for receiving the I/O operation (3416). The reader will understand that the host (3432) may track or have access to information describing other metrics (e.g., reliability of the metric, availability of the metric, throughput metric) that may be used alone or in combination to identify (3406) a particular storage system (3426) as a preferred storage system for receiving I/O operations (3416). Alternatively, the host (3402) may be configured to receive the identification of the preferred storage system from a system administrator as a configuration parameter, from the storage array itself, or in some other manner, such that identifying (3406) a particular storage system (3426) as a preferred storage system for receiving I/O operations (3416) may be performed simply by examining some configuration parameter or other configuration information stored within the host (3402).

[0496] The example method shown in FIG. 34 also includes, after identifying the preferred storage system (3426), issuing (3408) one or more I/O operations (3416) to the preferred storage system (3426) for the one or more I/O operations (3416) directed to the dataset (3418). In the example method shown in FIG. 34, the host (3402) may issue (3408) the one or more I/O operations (3416) directed to the dataset (3418) to the preferred storage system (3426) via one or more messages exchanged between the host (3402) and the preferred storage system (3426), for example, via a data communication path (812) between the host (3402) and the preferred storage system (3426).

[0497] For further explanation, FIG. 35 sets forth a flowchart illustrating an additional example method for managing connectivity to synchronously replicated data across storage systems (3424, 3426, 3428) in accordance with an embodiment of the present disclosure. While not shown in great detail, the storage systems (3424, 3426, 3428) illustrated in FIG. 35 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 35 may include the same components, fewer components, or additional components as the storage systems described above.

[0498] The example method illustrated in FIG. 35 is similar to the example method illustrated in FIG. 34 in that it also includes receiving (3404) I/O operations (3416) directed to a data set (3418) that is synchronously replicated across a plurality of storage systems (3424, 3426, 3428), identifying (3406) a particular storage system (3426) among the plurality of storage systems (3424, 3426, 3428) as a preferred storage system for receiving the I/O operations (3416), and, after identifying the preferred storage system (3426), issuing (3408) one or more I/O operations (3416) to the preferred storage system (3426) for the one or more I/O operations (3416) directed to the data set (3418).

[0499] The example method shown in FIG. 35 also includes determining (3502) respective response times for a plurality of storage systems (3424, 3246, 3428). In the example method shown in FIG. 35, the host (3402) determines (3502) respective response times for a plurality of storage systems (3424, 3246, 3428), for example, by determining the amount of time required by each of the storage systems (3424, 3426, 3428) to service similar I/O operations, by tracking the average amount of time required by each of the storage systems (3424, 3426, 3428) to service similar I/O operations, etc. In such an example, the host (3402) may track such information through the use of one or more internal clocks, by examining timestamps attached to one or more messages, or in some other manner.

[0500] In the example method shown in FIG. 35, identifying (3406) a particular storage system (3426) among the plurality of storage systems (3424, 3426, 3428) as a preferred storage system for receiving the I/O operation (3416) may include identifying (3504) a particular storage system among the plurality of storage systems (3424, 3426, 3428) as a preferred storage system for receiving the I/O operation (3416) according to respective response times for the plurality of storage systems among the plurality of storage systems (3424, 3426, 3428). In the example method shown in FIG. 35, the host (3402) may identify (3504) a particular storage system among the plurality of storage systems (3424, 3426, 3428) as a preferred storage system for receiving the I/O operation (3416) according to the respective response times for the plurality of storage systems among the plurality of storage systems (3424, 3426, 3428), for example, by selecting the storage system associated with the fastest response time as the preferred storage system, by selecting any storage system whose response time meets a predetermined quality of service threshold as the preferred storage system, or in some other manner.

[0501] The example method shown in FIG. 35 also includes detecting (3506) a change in response time for at least one of the storage systems (3424, 3426, 3428). In the example method shown in FIG. 35, the host (3402) detects (3506) the change in response time for at least one of the storage systems (3424, 3426, 3428) by determining that the average response time has deviated by more than a predetermined threshold amount as a result of performing additional tests on each of the storage systems, by detecting any interruption to the ability to exchange messages over a particular data communications link, or in some other manner.

[0502] The example method shown in FIG. 35 also includes selecting (3508) a different storage system of the plurality of storage systems (3424, 3426, 3428) as a preferred storage system according to a change in response time. In the example method shown in FIG. 35, the host (3402) selects (3508) a different storage system of the plurality of storage systems (3424, 3426, 3428) as a preferred storage system for receiving the I/O operation (3416), for example, by selecting the storage system associated with the fastest updated response time as the preferred storage system, by selecting any storage system whose updated response time meets a predetermined quality of service threshold as the preferred storage system, or in some other manner.

[0503] For further explanation, FIG. 36 sets forth a flowchart illustrating an additional example method for managing connectivity to synchronously replicated data across storage systems (3424, 3426, 3428) in accordance with an embodiment of the present disclosure. While not shown in great detail, the storage systems (3424, 3426, 3428) illustrated in FIG. 36 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage system illustrated in FIG. 36 may include the same components as the storage systems described above, fewer components, or additional components.

[0504] The example method illustrated in FIG. 36 is similar to the example method illustrated in FIG. 34 in that it also includes receiving (3404) I/O operations (3416) directed to a data set (3418) that is synchronously replicated across a plurality of storage systems (3424, 3426, 3428), identifying (3406) a particular storage system (3426) among the plurality of storage systems (3424, 3426, 3428) as a preferred storage system for receiving the I/O operations (3416), and, after identifying the preferred storage system (3426), issuing (3408) one or more I/O operations (3416) to the preferred storage system (3426) for the one or more I/O operations (3416) directed to the data set (3418).

36 also includes receiving (3602) an identification (3610) of a preferred storage system from one of the storage systems (3428). In the example method shown in FIG. 36, the host (3402) may receive (3602) the identification (3610) of the preferred storage system from one of the storage systems (3428) via one or more messages exchanged over a data communication path (3414) between the storage system (3428) and the host (3402). The host (3402) may maintain the identification (3610) of the preferred storage system, for example, as a configuration setting stored within the host. Thus, in other embodiments, the configuration setting may be set in a different manner (e.g., by a system administrator, by another software module running on the host), but the identification (3406) of a particular storage system (3426) of the plurality of storage systems (3424, 3426, 3428) as the preferred storage system for receiving the I/O operation (3416) may be performed in accordance with the configuration setting.

[0506] The example method shown in FIG. 36 also includes detecting (3604) that the host (3402) has moved from one location to another. In the example method shown in FIG. 36, the host (3402) may detect that it has moved from one location to another (3604), for example, by determining that the host (3402) has been connected to a new data communication interconnect, by detecting that the host (3402) is installed in a different position in a rack or in a new rack, or in some other manner. Location relative to a rack or within a data center, or based on the network topology of a data center or campus, may be an aspect of "locality" that may affect performance between a host and a particular storage system for a pod. In a single storage system that spans racks or is connected to multiple networks, locality may also apply to individual storage network adapters on individual storage systems in a pod.

36 also includes identifying (3608) a different storage system of the plurality of storage systems as a preferred storage system for receiving I/O operations directed to dataset (3418). In the example method shown in FIG. 36, host (3402) may identify (3608) a different storage system of the plurality of storage systems as a preferred storage system for receiving I/O operations directed to dataset (3418), for example, by re-measuring the response time associated with each of storage systems N and selecting the storage system exhibiting the fastest response time. In the example method shown in FIG. 36, identifying (3608) a different storage system of the plurality of storage systems as a preferred storage system for receiving I/O operations directed to dataset (3418) may be performed in response to detecting that host (3402) has moved.

[0508] The example method illustrated in FIG. 36 also includes detecting (3606) a configuration change to host (3402). In the example method illustrated in FIG. 36, host (3402) may detect (3606) a configuration change to host (3402), for example, by detecting that a different version of some software is installed on host (3402), by detecting that some hardware component in host (3402) has been changed or added, etc. In the example method illustrated in FIG. 36, identifying (3608) a different storage system of the plurality of storage systems as a preferred storage system for receiving I/O operations directed to dataset (3418) may instead be performed in response to detecting the configuration change to host (3402).

[0509] For further explanation, FIG. 37 sets forth a flowchart illustrating an example method for automated storage system configuration for an intermediation service according to some embodiments of the present disclosure. While not shown in great detail, the storage systems (3700A-3700N) illustrated in FIG. 37 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage systems (3700A-3700N) illustrated in FIG. 37 may include the same components as the storage systems described above, fewer components, or additional components.

[0510] In the following example, automated storage system configuration for an intermediation service may include determining whether a given storage system (3700A) among a collection of storage systems (3700A-3700N) is configured to request intermediation from an intermediation target or service - if the given storage system (3700A) is not configured to request intermediation, the given storage system (3700A) is configured to request or obtain an intermediation service handle from a pre-configured location before becoming operational. Such a determination of whether a storage system is configured to request intermediation may occur when the storage system is initially brought online or started up, and the storage system may be configured prior to shipping to request configuration information from a designated configuration service, such as configuration service (3751), which may operate as an independent computer system or within a third-party computing environment that provides configuration services.

[0511] Additionally, in some examples, determining whether a given storage system (3700A) among the storage systems (3700A-3700N) is configured to request mediation from the mediation service is performed in response to initiating dataset synchronous replication among the storage systems (3700A-3700N). In this example, in response to a storage system being added to a collection of storage systems that synchronously replicate a dataset or pod, a storage system that is currently a member of the pod may automatically transfer one or more handles for or related to the mediation service and mediation target(s) to the added storage system. In this manner, the first storage system receives the mediation target(s) from the configuration service, and each storage system added to the pod receives its mediation target(s) from the pod members, such that all member storage systems are configured to request mediation from the same mediation target(s).

[0512] As described above, a mediation service handle may be requested when a pod is created, when a pod is expanded, or when a pod is initially expanded so that a mediation service may be needed in the event of a future communication failure between storage systems within the pod. The mediation service handle may be a contact address on a wide area network and may be a cryptographically secured token that can be used to manage a pool of keys for handling mediation needs for the storage system cluster or pod. Alternatively, a first storage system in a pod may determine a secure handle to use with a well-known mediation service for use in a first or subsequent mediation competition, the handle being determined privately by that first storage system without any specific or necessary interaction with the mediation service, and the handle then being communicated to other storage systems already in the pod or as the pod.

[0513] The process of engaging an intermediary service in response to an error, such as a communication failure between storage systems—storage systems may be configured to store handles indicating contact addresses over a wide area network and cryptographically secure tokens that can be used to manage a pool of keys for the intermediary—is described in more detail in Application Serial No. 15/703,559, which is incorporated herein by reference in its entirety. Also described in Application Serial No. 15/703,559 is the use of various quorum protocols to determine which storage system in a collection of storage systems replicating a dataset should continue to service I/O requests directed to the dataset.

[0514] However, while this application Ser. No. 15/703,559 describes implementations for mediation and quorum protocols, the focus of this disclosure is automatic storage system configuration for the mediation service. In other words, a storage system may be set up before shipping to contact a configuration service and request a handle for the mediation service so that the storage system configures itself to handle communication failures through the mediation, and the storage system automatically transfers the mediation process as more storage systems are added to the pod—a user, such as an administrator, does not need to take any action to configure the storage system to perform the mediation.

[0515] As shown in FIG. 37, multiple storage systems (3700A-3700N) synchronously replicating dataset (3752) may communicate with each other via one or more networks (not shown) and with an intermediary service (3701)—the intermediary service (3701) may determine which storage system will continue to service the dataset in the event of a communication failure between the storage systems, if a storage system goes offline, or due to some other trigger event. Furthermore, in this example, in response to a storage system going offline, the configuration service (3751) may be automatically reachable by the storage system to request one or more handles to the intermediary service (3701) through the use of pre-configured contact addresses. In general, any number of storage systems may be part of an in-sync list that are synchronously replicating dataset (3752).

37 includes determining (3702) that a particular storage system (3700A) among the storage systems (3700A-3700N) is not configured to request intermediation from an intermediary target for intermediation between storage systems that synchronously replicate a dataset (3752). Determining (3702) that a particular storage system (3700A) among the storage systems (3700A-3700N) is not configured to request intermediation from an intermediary target may be implemented by the storage system (3700A) including a startup process within the controller, which may configure how a communication failure or an interruption related to a communication failure is handled. For example, the startup process of the controller (462) may check, for example, by reading a status flag or condition code indicating whether an intermediation handle has been determined, received, or requested. If the intermediation handler (3762) detects that an intermediation handle has already been configured, then this portion of the startup process is complete. However, if the intermediary handler (3762) detects that an intermediary handle has not been configured, the intermediary handler (3762) may proceed to determine such a handle or request an intermediary handle from the configuration service (3751) - the contact address for the intermediary service (3701) may be a system setting or system variable defined by the manufacturer before the storage system is shipped.

[0517] The example method shown in FIG. 37 also includes requesting (3704) configuration information indicating one or more service handles for the intermediary service (3701) by the particular storage system (3700A) from the configuration service (3751). Requesting (3704) configuration information indicating one or more service handles for the intermediary service (3701) from the configuration service (3751) by the particular storage system (3700A) may be performed by the intermediary handler (3762) accessing stored contact information for the configuration service (3751) and sending a request (3754) for one or more service handles for the intermediary service (3701).

[0518] The example method shown in FIG. 37 also includes configuring (3706) the intermediary handler to communicate with the intermediary service in response to detecting a communication failure with one of the storage systems (3700B-3700N) according to one or more service handles received from the configuration service (3751). Configuring (3706) the intermediary handler to communicate with the intermediary service in response to detecting a communication failure with one of the storage systems (3700B-3700N) according to one or more service handles received from the configuration service (3751) may be implemented by the intermediary handler (3762) defining a service handle or contact address for the intermediary service (3701) to be used when responding to the communication failure, and the service handle (3756) may be specified in a response message from the configuration service (3751) responding to the request (3754).

For additional explanation, Figure 38 sets forth a flowchart illustrating an example method for automated storage system configuration for an intermediation service in accordance with some embodiments of the present disclosure. The example method illustrated in Figure 38 is similar to the example method illustrated in Figure 37 in that it also includes determining (3702) that a particular storage system (3700A) among storage systems (3700A-3700N) is not configured to request intermediation from an intermediation target for intermediation between storage systems that synchronously replicate a dataset (3752), requesting (3704) configuration information indicating one or more service handles for the intermediation service (3701) by the particular storage system (3700A) from a configuration service (3751), and configuring (3706) an intermediation handler to communicate with the intermediation service in response to detecting a communication failure with one of the storage systems (3700B-3700N) according to the one or more service handles received from the configuration service (3751).
However, the example method shown in Figure 38 further includes automatically transferring 3802 one or more handles for the intermediary service to the storage system being added in response to adding the storage system to the storage systems (3700A-3700N) that synchronously replicate dataset 3752. Processes for adding a storage system to a pod are described herein such that the added storage system may become a member of the in-sync list. Given the described process for adding a storage system to a pod, transferring 3802 one or more handles for the intermediary service to the storage system being added may be implemented by further specifying the described process for sending a command to the system being added indicating that the service handle to be used by the intermediary handler is defined to be the service handle being sent in the command.

For further explanation, FIG. 39 illustrates a flowchart depicting an example method for automated storage system configuration of an intermediary service in accordance with some embodiments of the present disclosure. The example method illustrated in FIG. 39 is similar to the example method illustrated in FIG. 37 in that it also includes determining (3702) that a particular storage system (3700A) among storage systems (3700A-3700N) is not configured to request intermediation from an intermediary target for intermediation between storage systems that synchronously replicate a dataset (3752); requesting (3704) configuration information indicating one or more service handles for the intermediary service (3701) by the particular storage system (3700A) from a configuration service (3751); and configuring (3706) an intermediary handler to communicate with the intermediary service in response to detecting a communication failure with one of the storage systems (3700B-3700N) according to the one or more service handles received from the configuration service (3751).

[0521] However, the example method shown in FIG. 39 further includes providing (3902) a secure key or randomly selected key for mediation between the storage systems (3700A-3700N) by a particular storage system (3700A) to the intermediary service (3701). Providing (3902) a secure key or randomly selected key for mediation between the storage systems (3700A-3700N) by a particular storage system (3700A) to the intermediary service (3701) may be implemented as described in application Ser. No. 15/703,559, the entire contents of which are incorporated herein by reference.

[0522] For further explanation, FIG. 40 sets forth a flowchart illustrating an example method for automated storage system configuration for an intermediary service according to some embodiments of the present disclosure. While not shown in great detail, the storage systems (3700A-3700N) illustrated in FIG. 40 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3B, or any combination thereof. In practice, the storage systems (3700A-3700N) illustrated in FIG. 40 may include the same components as the storage systems described above, fewer components, or additional components.

[0523] In this example implementation of an automated storage system, instead of the storage system being pre-configured to request or obtain an intermediation service handle from a fixed, pre-configured location, such as a configuration service, before being shipped or installed, the storage system may be pre-configured with fixed contact information for the intermediation service. When creating a pod for a dataset, a given storage system may configure the use of the intermediation service to intermediate between storage systems in the pod by instantiating a key and providing the key to the intermediation service. Furthermore, when creating a pod for a dataset, a given storage system may configure other storage systems in the pod by providing the same generated key that can be provided to the intermediation service for each of the storage systems in the pod—e.g., as the storage system is added to the pod. Upon receiving the generated key, the other storage systems may use the generated key to request intermediation if they determine that intermediation is an appropriate response to a system failure.

[0524] In this example, the intermediary service may respond to a simple request to lock the use of a particular key; the intermediary service is not configured with the particular key prior to receiving it, and the intermediary service is not configured to receive the particular key from any particular storage system, pod, or customer. If a particular key is locked at the intermediary service by a given storage system, then other storage systems cannot lock that particular key at the intermediary service.

[0525] In general, keys may be generated or named using a cryptographically secure method to prevent exploit code, or hacking attempts, where an attacker could predict the value of a key and pre-lock it. Furthermore, a collection of storage systems within a cluster or pod may use a specific key for use in determining the next race for the cluster or pod, and may rotate a newly generated key to be used after each mediation attempt. In another example, after a storage system is added to a cluster or pod, one aspect of the pod configuration may include providing the added storage system with the current key to use for the next mediation request. For example, if some event triggers mediation, such as a storage device failure or network failure, one or more storage systems may then request a lock on the current key from the mediation service. This request for a lock on the current key may be successful for at most one requestor—if the response to the request is lost, the lock may be requested (or, in some embodiments, queried) again, and subsequent requests may be successful if the subsequent requests are identified by the mediation service as coming from the same requestor.

[0526] In some examples, after a single use of a particular brokering key, any remaining active storage systems in the cluster or pod may exchange for a new key for use in future brokering attempts. This exchange of a new key may also be performed if the previous brokering attempt was never confirmed by receiving a successful response from the brokering service—if the storage systems that were members at the time of the failure are to resume communication with each other.

[0527] In an alternative embodiment, a customer may explicitly configure an intermediary service, such as a service running on a customer-managed virtual machine. In yet another embodiment, there may be an intermediate step in configuring a cluster or pod whereby a storage system contacts an intermediary configuration service at an address to determine a service location address or other contact information that the cluster or pod will use for intermediation. In this embodiment, this intermediate step may be part of the core configuration process for the cluster or pod, whereby service location addresses or other contact information are exchanged between storage systems in the cluster or pod, and service location addresses or other contact information are sent to new storage systems as they are added to the cluster or pod. Such an embodiment may be useful in situations where a vendor tailors a specific intermediary service instance for a specific customer. Alternatively, such an embodiment may be used to explicitly locate an intermediary service based on terrain or to match a customer's location to a cloud service availability or assurance zone.

[0528] In some examples, with respect to an intermediary service, the intermediary service may be configured to be contacted by a storage system without customer interaction, and the intermediary service operates by locking on a particular instance of the intermediary in a cluster or pod and responding with success only if no conflicting requests have been made on the same lock, with success being used by one storage system in a cluster or pod as part of ensuring that it can safely resume service for the cluster or pod after detecting a failure that has isolated one storage system from at least one other storage system. The intermediary service may be a cloud-based service, such as a service that provides multiple front-end web servers that receive requests from one or more connecting clients, and the front-end web servers may advertise for a particular DNS hostname—which may be further visualized behind multiple network switches—that maps to any number of IP addresses. In some examples, multiple hostnames may be provided, and the front-end servers may be configured to scatter or distribute requests to the intermediary across multiple back-end servers. For example, in the case of an intermediation service, given a particular intermediation key associated with a particular lock request from a storage system, the particular key may then be hashed against any of multiple backend databases that enforce the lock for each received key. Furthermore, because multiple storage systems may compete to use the same key to intermediate for the same pod, the storage system may contact any of the frontend web servers—as long as they are associated with the same overall cloud-based intermediation service—and hash against the same backend database to enforce the lock. In this example, the backend database may be implemented as a distributed transactional database with appropriate guarantees (e.g., DynamoDB™, among others); as a highly available database server on shared storage; as a synchronously replicated database server with appropriate mechanisms for high availability and data redundancy; or using an object storage model with appropriate guaranteed conditional storage primitives; or the backend database may be implemented with any of a variety of other technologies.

[0529] Additionally, in some examples, the intermediary service may be implemented to ensure that each key—whether it be a cluster key, a pod key, a customer key, or a key associated with another type of domain—is securely isolated from other keys, clusters, pods, customers, or other types of domains, depending on the implementation. This can ensure, for example, an effective multi-tenancy model for the intermediary service.

40 includes determining (4002) to configure one or more of the storage systems (3700A-3700N) by a particular storage system (3700A) from an intermediary service (3701) for intermediation between the storage systems (3700A-3700N) that synchronously replicate a data set (3752). Determining (4002) to configure one or more of the storage systems (3700A-3700N) by a particular storage system (3700A) from an intermediary service (3701) for intermediation between the storage systems (3700A-3700N) that synchronously replicate a data set (3752) by a particular storage system (3700A) from the intermediary service (3701) for intermediation between the storage systems (3700A-3700N) that synchronously replicate a data set (3752) may be implemented using different techniques. In one example, determining to configure a storage system (4002) may be performed based on expanding a cluster or pod to include one or more additional storage systems, and as new storage systems are added to the cluster or pod, each new storage system is then provided with a key for mediation that each other existing storage system is configured to use for mediation. In another example, determining to configure a storage system (4002) may be performed by maintaining, for each storage system that is a member of a pod that synchronously replicates the dataset (3752), metadata describing whether a given storage system has been provided with an instance of a key, and the initial state of the metadata may indicate that the storage system does not have a key. For example, if there is no current consensus on a key, the particular storage system (3700A) may generate an instance of a key as described above, using cryptographic techniques that generate keys that may not be computationally feasible to predict. Furthermore, the particular storage system (3700A) may determine, during startup or periodically, that a key has not been generated—i.e., there is no current consensus on a key value—and may generate an instance of a key accordingly. Additionally, in some examples, the intermediary service may be part of the replicated content, including metadata, of the synchronized data sets (3752) among the storage systems that are in sync with each other. A quorum policy may then be used between the storage systems as part of an overall set of algorithms to determine which storage systems will remain in sync with each other. In this way, the storage systems (3700A-3700N) may use the same instance of a key when the intermediary service is used.

40 also includes providing (4004) an instance of key (4052) to one or more of storage systems (3700A-3700N) to request intermediation from the intermediary service (3701). Providing (4004) an instance of key (4052) to one or more of storage systems (3700A-3700N) to request intermediation from the intermediary service (3701) may be implemented by transmitting, from the particular storage system (3700A), an instance of key to each of two or more storage systems determined not to be configured to request intermediation from the intermediary service, using one or more network ports and across one or more communication networks. In some cases, providing (4004) an instance of the key to one or more storage systems (3700B-3700N) may be performed by a particular storage system (3700A) of the storage systems (3700A-3700N) in response to determining (4002) that one or more of the storage systems is not configured to request intermediation from the intermediation service (3701) for intermediation between the storage systems (3700A-3700N) that synchronously replicate the dataset (3752).

[0532] In some embodiments, under normal circumstances, a pod may start with an intermediary key and will transmit the intermediary key as the pod expands to include additional storage systems—a particular storage system, which in some cases may be considered the “leader,” will transmit a new intermediary key required for subsequent intermediation competitions in response to a key being used once. In other words, in contrast to the above embodiments, in some embodiments, instead of explicitly determining whether a storage system is configured to use the intermediary key, the storage system may have access to the intermediary key based on the creation of the pod as part of the pod's creation—in response to a storage system being added to the pod, the intermediary key is distributed to the new storage system as part of the synchronization of data and metadata across the storage systems in the pod. Further, as described above, in this embodiment, in response to using or attempting to use the intermediary key, new keys may be generated and distributed to storage systems in the pod. In this manner, storage systems in the pod may be configured to use the same intermediary key generated in response to the creation of the pod, expand the pod to include additional storage systems, and use or attempt to use the intermediary key. As described above, configuring to use an intermediary key may include specifying that the current key available to the intermediary handler is the intermediary key that is synchronized across the pod's storage systems.

40 also includes providing (4006) an instance of key (4052) to intermediary service (3701), where the intermediary service provides the intermediary service to a given storage system that provides the instance of key (4052). Providing (4006) the instance of key (4052) to intermediary service (3701), where the intermediary service provides the intermediary service to a given storage system that provides the instance of key (4052), may be implemented by transmitting the instance of key (4052) from a particular storage system (3700A) to intermediary service (3701) using one or more network ports and across one or more communication networks. In some cases, providing (4006) an instance of the key (4052) to the intermediary service (3701), in which the intermediary service provides the intermediary service to a given storage system that provides the instance of the key (4052), may be performed in response to determining (4002) by a particular storage system (3700A) among the storage systems (3700A-3700N) that one or more of the storage systems is not configured to request intermediation from the intermediary service (3701) for intermediation between the storage systems (3700A-3700N) that synchronously replicate the dataset (3752).

[0534] The example method shown in FIG. 40 also includes configuring (708) the intermediary handler (3762) to provide an instance of the intermediary service's key to the intermediary service in response to detecting a communication failure with at least one of the storage systems (3700B-3700N). Configuring (4006) the intermediary handler (3762) to provide an instance of the intermediary service's key to the intermediary service in response to detecting a communication failure with at least one of the storage systems (3700B-3700N) may be implemented by the intermediary handler (3762) defining the current key to be the instance of the generated key provided to the intermediary service (3701) when responding to the communication failure.

[0535] For further explanation, FIG. 41 illustrates a diagram of metadata representations that may be implemented as a structured collection of metadata objects that together may represent a logical volume or portion of a logical volume of storage data, in accordance with some embodiments of the present disclosure. Metadata representations 4150, 4154, and 4160 may be stored within storage system (4106), and one or more metadata representations may be generated and maintained for each of multiple storage objects, e.g., volumes or portions of volumes, stored within storage system (4106).

[0536] While other types of structured collections of metadata objects are possible, in this example the metadata representation may be structured as a directed acyclic graph (DAG) of nodes, and the DAG may be structured and balanced according to a variety of methods to maintain efficient access to any given node. For example, the DAG for the metadata representation may be defined as a type of B-tree and balanced accordingly in response to changes to the structure of the metadata representation, which may occur in response to or in addition to changes to the underlying data represented by the metadata representation. While in this example there are only two levels for simplicity, in other examples the metadata representation may span multiple levels and include hundreds or thousands of nodes, with each node including any number of links to other nodes.

[0537] Further, in this example, the leaves of the metadata representation may include pointers to data stored for the volume or portion of the volume, and the logical address, i.e., volume and offset, may be used to identify and navigate through the metadata representation to reach one or more leaf nodes that reference the stored data corresponding to the logical address. For example, volume (4152) may be represented by metadata representation (4150) that includes metadata object nodes (4152, 4152A-4152N), and the leaf nodes (4152A-4152N) include pointers to respective data objects (4153A-4153N, 4157). A data object may be a unit of data of any size within storage system (4106). For example, data objects (4153A-4153N, 4157) may each be a logical extent, and the logical extent may be some specified size, such as 1 MB, 4 MB, etc., or some other size.

[0538] In this example, snapshot (4156) may be created as a snapshot of a stored object, in this case volume (4152), and at the time snapshot (4156) is created, metadata representation (4154) for snapshot (4156) includes all of the metadata objects for metadata representation (4150) for volume (4152). Furthermore, in response to the creation of snapshot (4156), metadata representation (4154) may be designated as read-only. However, volume (4152), which shares the metadata representation, may continue to be modified, and at the moment the snapshot is created, as modifications are made to the data corresponding to volume (4152), the metadata for volume (4152) and snapshot (4156) are identical; in response to the modifications, the metadata representation of volume (4152) and snapshot (4156) diverge and become different.

[0539] For example, given a metadata representation (4150) for representing a volume (4152) and a metadata representation (4154) for representing a snapshot (4156), the storage system (4106) may receive an I/O operation to write data that will ultimately be stored in a particular data object (4153B), the data object (4153B) being pointed to by a leaf node pointer (4152B), which is part of the metadata representations (4150, 4154). In response to the write operation, the read-only data objects (4153A-4153N) referenced by the metadata representation (4154) may remain unchanged, and the pointers (4152B) may also remain unchanged. However, metadata representation 4150 representing current volume 4152 is modified to include a new data object to hold the data written by the write operation, and the modified metadata representation is shown in metadata representation 4160. Furthermore, the write operation may be directed to only a portion of data object 4153B, and as a result, new data object 4157 may include a copy of the previous contents of data object 4153B in addition to the payload for the write operation.

[0540] In this example, as part of processing the write operation, the metadata representation (4160) for volume (4152) is modified to remove the existing metadata object pointer (4152B) and to include a new metadata object pointer (4158), which is configured to point to a new data object (4157), which stores the data written by the write operation. Furthermore—with the exception of metadata object pointer (4152B), which referenced the target data object, the metadata representation (4160) for volume (4152) continues to include all of the metadata objects contained in the previous metadata representation (4150), and metadata object pointer (4152B) continues to reference the read-only data object (4153B) that may have been overwritten.

[0541] In this way, using the metadata representation, a volume or portion of a volume may be considered to be snapshotted or copied by creating a metadata object and without actual duplication of data objects—duplication of data objects may be deferred until a write operation is directed to one of the read-only data objects referenced by the metadata representation.

[0542] In other words, the advantage of using a metadata representation to represent a volume is that a snapshot or copy of the volume may be created and accessible at a certain older time, specifically within the time it takes to create a metadata object for the snapshot or copy and to create a reference for the snapshot or copy metadata object to the existing metadata representation of the volume being snapshotted or copied.

[0543] By way of example, virtualized copy-by-reference may utilize metadata representations in a manner similar to the use of metadata representations when creating a snapshot of a volume—the metadata representation for virtualized copy-by-reference may often correspond to a portion of the metadata representation for the entire volume. An example implementation of virtualized copy-by-reference may be in the context of a virtualized storage system, where multiple block ranges within and between volumes may reference a unified copy of the stored data. In such a virtualized storage system, the metadata described above may be used to handle the relationship between virtual addresses, i.e., logical addresses, and physical addresses, i.e., real addresses—in other words, the metadata representation of stored data enables a virtualized storage system that may be considered flash-friendly in that it reduces or minimizes wear on flash memory.

[0544] In some examples, logical extents may be combined in a variety of ways, including as simple aggregations or as logically related address ranges within some larger logical extent formed as a collection of logical extent references. These larger associations may be given various types of logical extent identities and may be further combined into even larger logical extents or aggregations. Copy-on-write status may be applied at various tiers and in various ways depending on the implementation. For example, copy-on-write status applied to one logical collection of extents may cause the copied collection to hold references to unchanged logical extents, creating a copy-on-write logical extent (by copying references to any unchanged stored data blocks as needed) when only a portion of the copy-on-write logical collection is modified.

[0545] Deduplication, volume snapshots, or block range snapshots may be implemented in this model by a combination of referencing stored data blocks, referencing logical extents, or marking logical extents (or identified collections of logical extents) as copy-on-write.

[0546] Furthermore, with flash storage systems, stored data blocks may be organized and grouped together in a variety of ways. Because aggregates are written out into pages that are part of larger erase blocks, the eventual garbage collection of deleted or replaced stored data blocks may involve moving the contents stored in some number of pages elsewhere so that the entire erase block is erased and ready for reuse. This process of selecting physical flash pages, ultimately migrating and garbage collecting them, and then erasing flash erase blocks for reuse may or may not be coordinated, driven, or performed by aspects of the storage system that also handle logical extents, deduplication, compression, snapshots, virtual copies, or other storage system functions. The coordinated or driven process for selecting pages, migrating pages, garbage collecting, and erasing erase blocks may further take into account various characteristics of flash memory device cells, pages, and erase blocks, such as, for example, number of uses, aging projections, adjustments for voltage levels or number of retries previously required to recover stored data, etc. They may also take into account analysis and prediction across all flash memory in the storage system.

[0547] Continuing with this example, in which a storage system may be implemented based on a directed acyclic graph containing logical extents, logical extents can be classified into two types: leaf logical extents, which reference some amount of stored data in some way, and composite logical extents, which reference other leaf or composite logical extents.

[0548] Leaf extents can reference data in a variety of ways. A leaf extent may directly point to a single range of stored data (e.g., 64 kilobytes of data), or it may be a collection of references to stored data (e.g., a 1 megabyte "range" of contents that maps some number of virtual blocks associated with a range to physically stored blocks). In the latter case, these blocks may be referenced using some identity, and some blocks within the extent's range may not map to anything. Also, in the latter case, these block references need not be unique, allowing multiple mappings from virtual blocks in some number of logical extents within and across some number of volumes to map to the same physically stored block. Instead of stored block references, logical extents could encode simple patterns. For example, blocks that are strings of identical bytes could simply encode that the block is a repeated pattern of identical bytes.

[0549] A composite logical extent may be a logical range of content having some virtual size that includes multiple maps, each mapping a logical range of content from a subrange of the composite logical extent to an underlying leaf logical extent or composite logical extent. Transforming a request related to content in a composite logical extent then involves taking the content range for the request within the context of the composite logical extent, determining which underlying leaf logical extent or composite logical extent the request maps to, and transforming the request to apply to the appropriate range of content within those underlying leaf logical extents or composite logical extents.

[0550] Volumes, files, or other types of storage objects can be described as composite logical extents. Therefore, these represented storage objects can be organized using this extent model.

[0551] Depending on the implementation, a leaf logical extent or a composite logical extent may be referenced from multiple other composite logical extents, essentially allowing for inexpensive duplication of larger collections of content within and across volumes. Thus, logical extents may essentially be arranged in an acyclic graph of references, each terminating in a leaf logical extent. This can be used to create copies of volumes, to create snapshots of volumes, or as part of virtual range copy support within and between volumes as part of an EXTENDED COPY or similar type of operation.

[0552] Implementations may provide each logical extent with an identity that can be used to name it. This simplifies referencing, as the reference within a composite logical extent will be a list containing the logical extent identifier and the logical subrange corresponding to each such logical extent identity. Also, within a logical extent, each stored data block reference may be based on some identity used to name it.

[0553] To support these overlapping uses of extents, an additional feature can be added: copy-on-write logical extents. When a modification operation affects a copy-on-write leaf logical extent or composite logical extent, the logical extent is copied, and the copy is a new reference and possibly has a new identity (depending on the implementation). The copy preserves all references or identities related to the underlying leaf logical extent or composite logical extent, regardless of any modifications resulting from the modification operation. For example, a WRITE, WRITE SAME, XDWRITERREAD, XPWRITE, or COMPARE AND WRITE request may store new blocks in the storage system (or use deduplication techniques to identify existing stored blocks) and modify the corresponding leaf logical extent to reference or store identities for the new set of blocks, possibly replacing references and stored identities for the previous set of blocks. Alternatively, an UNMAP request may modify the leaf logical extent to remove one or more block references. In both types of cases, a leaf logical extent is modified. If the leaf logical extent is copy-on-write, then a new leaf logical extent is created, formed by copying unaffected block references from the old extent and then replacing or deleting block references based on the modification operation.

[0554] The composite logical extent used to locate the leaf logical extent may then be modified to store a new leaf logical extent reference or identity associated with the copied and modified leaf logical extent as a replacement for the previous leaf logical extent. If the composite logical extent is copy-on-write, then a new composite logical extent is created with the new reference or identity, any unaffected references or identities to its underlying logical extents are copied to the new composite logical extent, and the previous leaf logical extent reference or identity is replaced with the new logical extent reference or identity.

[0555] This process continues from the referenced extent back to the referencing composite extents, based on a search path through the acyclic graph used to process the modification operation, until all copy-on-write logical extents are copied, modified, and replaced.

[0556] These copied leaf logical extents and composite logical extents may then exclude the copy-on-write feature, whereby additional modifications do not result in additional copies. For example, the first time some underlying logical extent in a copy-on-write "parent" composite extent is modified, that underlying logical extent is copied and modified, and the copy then has a new identity written into the copied, replaced instance of the parent composite logical extent. However, the second time some other underlying logical extent is copied and modified, a new identity for the other underlying logical extent copy is written into the parent composite logical extent, and the parent may then be modified in place without additional copies, replacing the necessary references to the parent composite logical extent.

[0557] Modification operations to new areas of a volume or composite logical extent for which there is no current leaf logical extent may create a new leaf logical extent to store the results of those modifications. If that new logical extent is referenced from an existing copy-on-write logical extent, then that existing copy-on-write composite logical extent references the new logical content, resulting in another sequence of copy, modify, and replace operations similar to the sequence for modifying an existing leaf logical extent.

[0558] If a parent composite logical extent cannot grow large enough (based on the implementation) to cover the associated address range, including the new leaf logical extent to create for the new modification operation, then the parent composite logical extent may be copied to two or more new composite logical extents that are in turn referenced from a single "grandchild" composite logical extent that is itself a new reference or new identity. If that grandchild logical extent is found through another composite logical extent that is itself copy-on-write, then that other composite logical extent is copied, modified, and replaced in a manner similar to that described in the preceding paragraph. This copy-on-write model can be used as part of implementing snapshots, volume copies, and virtual volume address range copies in storage system implementations based on these directed acyclic graphs of logical extents. To create a snapshot as a read-only copy of an otherwise writable volume, the graph of logical extents associated with the volume is marked copy-on-write, and a reference to the original composite logical extent is maintained by the snapshot. Modification operations on the volume then create logical extent copies as needed, resulting in a volume that stores the results of those modification operations and a snapshot that preserves the original contents. Volume copies are similar, except that both the original and copied volumes modify content, resulting in their own copied logical extent graphs and subgraphs.

[0559] Virtual volume address range copies may operate by copying block references within and between leaf logical extents (which does not itself involve using copy-on-write techniques unless changes to the block references modify the copy-on-write leaf logical extents). Alternatively, virtual volume address range copies may duplicate references to leaf logical extents or composite logical extents, which works well for volume address range copies of larger address ranges. Furthermore, this allows the graph to be a directed acyclic graph of references rather than simply a reference tree. Copy-on-write techniques associated with duplicated logical extent references can be used to ensure that modification operations on the source or target of a virtual address range copy result in the creation of new logical extents to store those modifications without affecting the target or source that share the same logical extent immediately after the volume address range copy operation.

[0560] Input/output operations for a pod may also be performed based on replicating a directed acyclic graph of logical extents. For example, each storage system in a pod may implement a private graph of logical extents, such that the graph on one storage system for the pod has no particular relationship to the graph on any second storage system for the pod. However, there is value in synchronizing the graphs between storage systems within a pod. This may be useful for resynchronization and for coordinating features such as asynchronous replication or snapshot-based replication to remote storage systems. Furthermore, it may also be useful for handling snapshot distribution and reducing some overhead for copy relationship processing. In such a model, keeping the contents of a pod in sync across all synchronizing storage systems for the pod is essentially the same as keeping the graphs of leaf logical extents and composite logical extents in sync for all volumes across all synchronizing storage systems for the pod and ensuring that the contents of all logical extents are in sync. To be in sync, matching leaf logical extents and composite logical extents should either have the same identity or should have mappable identities. Mapping might involve some set of intermediate mapping tables, or it might involve some other type of identity translation. In some cases, the identities of blocks mapped by leaf logical extents could also be kept in sync.

[0561] In a leader-and-follower based pod implementation with a single leader per pod, the leader may be responsible for determining any changes to the logical extent graph. When a new leaf logical extent or composite logical extent is created, it may be given an identity. When an existing leaf logical extent or composite logical extent is copied to form a new logical extent with modifications, the new logical extent may be described as a copy of the previous logical extent with some set of modifications. When an existing logical extent is split, the split may be described along with the resulting new identity. When a logical extent is referenced as a base logical extent from some additional composite logical extent, the reference may be described as a change to the composite logical extent to reference that base logical extent.

[0562] Modification operations in a pod thus include distributing descriptions of modifications to the logical extent graph (when new logical extents are created to extend content, or when logical extents are copied, modified, or replaced to handle copy-on-write conditions related to snapshots, volume copies, and volume address range copies), and distributing descriptions and contents for modifications to the content of leaf logical extents. An additional advantage of using metadata in the form of a directed acyclic graph, as described above, is that I/O operations that modify stored data in physical storage can be effected at the user level—without modifying the data stored in physical storage—by modifying the metadata corresponding to the data stored in physical storage. In disclosed embodiments of a storage system in which the physical storage may be a solid-state drive, wear associated with modifications to flash memory can be avoided or reduced due to I/O operations being effected by modifying metadata representing the data targeted by the I/O operation, instead of through a read, erase, or write to flash memory. Furthermore, in a virtualized storage system, the metadata described above may be used to handle the relationship between virtual addresses, i.e., logical addresses, and physical addresses, i.e., real addresses - in other words, the metadata representation of stored data enables a virtualized storage system that may be considered flash-friendly in that it reduces or minimizes wear on flash memory.

[0563] The leader storage system may perform its own local operations to implement these descriptions in the context of its local copy of the pod dataset and the metadata of its local storage system. Additionally, synchronizing followers may perform their own separate local operations to implement these descriptions in the context of their separate local copies of the pod dataset and the metadata of their separate local storage systems. When both the leader and follower operations are complete, the result is a compatible graph of logical extents with compatible leaf logical extent contents. These graphs of logical extents then become a kind of "common metadata," as described in the example above. This common metadata may be described as dependencies between modification operations and required common metadata. The transformation into a graph can be described as a set of predicates, or separate operations within more predicates, that may describe relationships, such as dependencies, along with one or more other operations. In other words, interdependencies between operations may be described as a set of precursors on which an operation depends in some way, and the set of precursors may be viewed as predicates that must be true for the operation to complete. A more complete description of predicates may be found in Application Serial No. 15/696,418, which is incorporated herein by reference in its entirety. Instead, each modification operation that relies on the same particular graph transformation not yet known to be completed across the entire pod may include portions of any graph transformations it relies on. Processing an operation description that identifies a "new" leaf or composite logical extent that already exists can avoid creating a new logical extent because that portion was already processed in the processing of some earlier operation, and instead implement only the portion of the operation processing that modifies the contents of the leaf or composite logical extent. It is the reader's responsibility to ensure that the transformations are compatible with each other. For example, a pod may begin with two writes receiving the pod. The first write replaces composite logical extent A with a copy created as composite logical extent B, replaces leaf logical extent C with a copy created as leaf logical extent D and modifications to store the contents for the second write, and then writes leaf logical extent D to composite logical extent B. Meanwhile, a second write implies the same copy of composite logical extent A and its replacement with composite logical extent B, but copies a different leaf logical extent E and replaces it with logical extent F that is modified to store the content of the second write, and writes logical extent F to logical content B. In that case, a description of the first write may include the replacement of A with B and the replacement of C with D, and writing D to composite logical extent B, and writing the content of the first write to leaf extent B, and a description of the second write may include the replacement of A with B and the replacement of E with F, and writing F to composite logical extent B, with the content of the second write written to leaf extent F. The leader or any follower can then process the first or second write in any order, with the end result being that B copies and replaces A, D copies and replaces C, F copies and replaces E, and D and F are written to the composite logical extent B. The second copy of A to form B can be avoided by recognizing that B already exists. In this way, the leader may ensure that compatible common metadata is maintained for the logical extent graph across synchronized storage systems for the pod.

[0564] Given an implementation of a storage system that uses a directed acyclic graph of logical extents, recovery of a pod may be performed based on the replicated directed acyclic graph of logical extents. Specifically, in this example, recovery at a pod may be based on the replicated extent graph and then involve not only recovering the contents of leaf logical extents but also restoring consistency to these graphs. In this implementation of recovery, operations may include querying not only all leaf logical extent content modifications that are not known to be completed across all storage systems for the pod, but also graph transformations that are not known to be completed across all synchronizing storage systems for the pod. Such queries may be based on operations since some coordinated checkpoint, or simply operations that are not known to be completed; each storage system maintains a list of operations during normal operation that have not yet signaled as completed. In this example, the graph transformations are simple. Graph transformations create new things, copy old things to new things, and may copy old things to two or more separate new things; they modify composite extents to modify their references to other extents. Any stored operation description found on any in-sync storage system that creates or replaces any logical extent may be copied and executed on any other storage system that does not already have that logical extent. As long as the involved leaf logical extent or composite logical extent has been properly recovered, operations that describe modifications to a leaf logical extent or composite logical extent can apply those modifications to any in-sync storage system that has not yet applied them.

[0565] In another example, as an alternative to using a logical extent graph, storage may be implemented based on a replicated content-addressable store. In a content-addressable store, for each block of data (e.g., every 512 bytes, 4096 bytes, 8192 bytes, or even every 16384 bytes), a unique hash value (sometimes called a fingerprint) is calculated based on the block contents, allowing a volume or volume extent range to be described as a list of references to blocks with particular hash values. In a synchronously replicated storage system implementation based on references to blocks with the same hash value, replication might involve a first storage system receiving blocks, calculating fingerprints for those blocks, identifying block references for those fingerprints, and delivering the changes to one or more additional storage systems as updates to the mapping of volume blocks to the referenced blocks. If a block is found to be already stored by the first storage system, that storage system can use that reference to name the reference on each of the additional storage systems (either because the references use the same hash value, or because the identifiers for the references are identical or easily mappable). Alternatively, if a block cannot be found by a first storage system, then the contents of the first storage system, along with the hash value or identity associated with the block contents, may be delivered to the other storage system as part of the operation description. Furthermore, the volume description of each synchronizing storage system is then updated with the new block reference. Recovery in such a store may then include comparing recently updated block references for the volume. If block references differ between different synchronizing storage systems for a pod, then one version of each reference may be copied to the other storage system to make them consistent. If a block reference on one system does not exist, then it is copied from some storage system that stores the block for that reference. Virtual copy operations may be supported in such block or hash reference stores by copying the reference as part of performing the virtual copy operation.

[0566] For further explanation, FIG. 42A sets forth a flowchart illustrating an example method for synchronizing metadata between storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage system (4200A) shown in FIG. 42A may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3C, or any combination thereof. In practice, the storage system (4200A) shown in FIG. 42A may include the same components as the storage systems described above, fewer components, or additional components.

[0567] As described above, metadata may be synchronized between storage systems that are synchronously replicating a data set. Such metadata, sometimes referred to as common metadata or shared metadata, is stored by a storage system on behalf of a pod related to mappings of segments of content stored in the pod to virtual addresses in storage objects within the pod, and information related to these mappings is synchronized between member storage systems for the pod to ensure correct behavior—or better performance—for storage operations related to the pod. In some examples, a storage object may implement a volume or snapshot. The synchronized metadata may include (a) information for keeping volume content mappings synchronized between system storage within the pod, (b) tracking data for recovery checkpoints or for ongoing write operations, and (c) information related to the delivery of data and mapping information to remote storage systems for asynchronous or periodic replication.

[0568] Information for keeping content mappings synchronized between storage systems within a pod may enable efficient creation of snapshots, which in turn may enable subsequent updates, snapshot copies, or snapshot deletions to be performed efficiently and consistently across pod member storage systems.

[0569] Tracking data for recovery checks or for ongoing write operations may enable efficient collision recovery and efficient detection of content or volume mappings that may have been partially or fully applied on individual storage systems for a pod, but may not have been fully applied on other storage systems for the pod.

[0570] Information related to the delivery of data and mapping information to a remote storage system for asynchronous or periodic replication may enable two or more member storage systems for a pod to act as servers for replicated pod content with minimal concern for addressing discrepancies in mapping and identifying metadata used to drive asynchronous or periodic replication.

[0571] In some examples, the shared metadata may include an identifier for a named group or group representation of one or more volumes or one or more storage objects that are a subset of the entire synchronously replicated dataset for a pod—such a set of dataset volumes or storage objects may be referred to as a consistency group. A consistency group may be defined to specify a set of dataset volumes or storage systems to be used for consistent snapshots, asynchronous replication, or periodic replication. In some examples, a consistency group may be dynamically computed, for example, by including all volumes connected to a particular set of hosts or host network ports, or connected to a particular set of applications, virtual machines, or containers, which may be running on external server systems or on one or more of the storage systems that are members of the pod. In other examples, a consistency group may be defined according to a user selection of a type of data or set of data, or a consistency group specification similar to dynamic computation, where a user may specify, for example, through a command or management console, that a particular or named consistency group be created to include all volumes connected to a particular set of hosts or host network ports, or to include data for a particular set of applications, virtual machines, or containers.

[0572] In examples using consistency groups, a first consistency group snapshot of a consistency group may include a first set of snapshots for all volumes or other storage objects that are members of the consistency group at the time of the first dataset snapshot, and a second consistency group snapshot of the same consistency group may include a second set of snapshots for volumes or other storage objects that are members of the consistency group at the time of the second dataset snapshot. In other examples, dataset snapshots may be stored asynchronously on one or more target storage systems. Similarly, asynchronous replication of a consistency group may account for dynamic changes to member volumes and other storage objects of a consistency group, and a consistency group snapshot of a consistency group at either the source or target of an asynchronous replication link includes the volumes and other storage objects that are members in the consistency group relationship at the time the dataset snapshot pertains. In the case of a target of an asynchronous replication connection, the time to which a dataset snapshot pertains depends on the sender's dynamic dataset when it was received and was in process at the time of the consistency group snapshot at the target. For example, if the target of asynchronous replication is, say, 2000 operations behind, some of those operations being consistency group member changes, with a first set of such changes being more than 2000 operations ago to the source, and a second set of changes being within the last 2000, then a consistency group snapshot at that point in time at the target will account for the first set of member changes and not the second set of changes. Other uses of the target of asynchronous replication may similarly account for the point-in-time nature of the dataset for the consistency group in determining the volumes or other storage objects (and their contents) for their use. For example, in the same case where the asynchronous replication is 2000 operations behind, use of the target for disaster recovery failover could start with a dataset containing the volumes and other storage objects (and their contents) as they were at the source 2000 operations ago. For the purposes of this description, operations occurring concurrently at the source (e.g., writes, creation or deletion of storage objects, changes to properties that affect the inclusion or exclusion of other data from volumes or other storage objects or consistency data, or other operations that were in progress and not signaled as completed at the same point in time) may not have a single, well-defined ordering, so the counts of operations need only represent some plausible ordering based on any possible ordering of operations occurring concurrently at the source.

[0573] As another example of using consistency groups, in the case of periodic replication based on replicating consistency group snapshots, each replicated consistency group snapshot would include the volumes and other storage objects as they were at the time each consistency group snapshot was created at the source. Ensuring that consistency group membership is kept consistent through the use of common, or shared, metadata ensures that a failure—or a failure that may cause the source of a replication or the system that creates the data set snapshot to switch from one storage system to another within a pod—does not result in the loss of information needed to properly process those consistency group snapshots or consistency group replicas. Furthermore, this type of processing may allow multiple storage systems that are members of a pod to simultaneously serve as source systems for asynchronous or periodic replication.

[0574] Additionally, the synchronized metadata describing the mapping of segments to storage objects is not limited to the mapping itself, but may include additional information such as, for example, a sequence number (or some other value for identifying the stored data), a timestamp, a volume/snapshot relationship, a checkpoint identity, a tree or graph defining a hierarchy, or a directed graph of the mapping relationship, among other storage system information.

[0575] As shown in FIG. 42A, multiple storage systems (4200A-4200N) synchronously replicating a dataset (4258) may communicate with each other storage system (4200B-4200N) in the in-sync list for the pod—the storage systems may exchange metadata describing I/O operations to perform and updates to be made to each local metadata representation of the dataset (4258) stored on the individual storage system. Additionally, each storage system (4200A, 4200B, ... 4200N) may store a respective version of the storage object (4256, 4260...4262).

[0576] The example method shown in FIG. 42A includes receiving (4202) at a first storage system (4200A) of storage systems (4200A-4200N) an I/O operation (4252) directed to a dataset (4258). Receiving (4202) at a first storage system (4200A) of storage systems (4200A-4200N) the I/O operation (4252) directed to the dataset (4258) may be performed using one or more communication protocols for transporting packets or data across a network, such as, for example, a storage area network (158), the Internet, or any computer network through which a host computer (4251) may communicate with storage system (4200A). In this example, storage system (4200A) may receive an I/O operation (4252) received at a network port, such as a SCSI port, where the I/O operation (4252) is a write command directed to a memory location that is part of a dataset (4258) that is synchronously replicated across storage systems (4200A-4200N) in the pod.

[0577] The example method shown in FIG. 42A also includes determining (4204) metadata updates (4254) describing mappings of segments of content to virtual addresses in a storage object (4256) in accordance with the I/O operation (4252), where the storage object (4256) includes a dataset (4258). Determining (4204) metadata updates (4254) describing the mapping of segments of content to virtual addresses in a storage object (4256) according to an I/O operation (4252), where the determination that the storage object (4256) contains a dataset (4258) may be performed by determining or identifying information as described above regarding the metadata content to be synchronized across the pod's storage systems (4200A-4200N); information from the I/O operation (4252) may also be included in the metadata updates (4254), such logic, i.e., virtual addresses, payload size, and other information such as deduplication information describing how the I/O operation (4252) payload is contained or incorporated with respect to data previously stored in the dataset (4258).

[0578] The example method shown in FIG. 42A also includes synchronizing (4206) the metadata at a second storage system (4200B) of the storage systems (4200A-4200N) by sending a metadata update (4254) to the second storage system (4200B) to update a metadata representation on the second storage system according to the metadata update (4254). Synchronizing (4206) the metadata at a second storage system (4200B) of the storage systems (4200A-4200N) by sending a metadata update (4254) to the second storage system (4200B) to update a metadata representation on the second storage system according to the metadata update (4254) may be performed. This may be implemented by sending metadata updates (4254) to each of the other storage systems (4200B through 4200N) in the pod using one or more network ports and across one or more communication networks (not shown), which may receive the metadata updates (4254) to update their respective local metadata representations of the synchronized dataset (4258). After each storage system (4200B through 4200N) receives and processes the metadata updates (4254), the metadata corresponding to the synchronized dataset (4258) is synchronized across all systems.

For further explanation, FIG. 42B illustrates a flowchart depicting an example method for synchronizing metadata between storage systems synchronously replicating a dataset in accordance with some embodiments of the present disclosure. The example method depicted in FIG. 42B includes receiving (4202), at a first storage system (4200A) of storage systems (4200A-4200N), an I/O operation (4252) directed to a dataset (4258); and determining (4204), in accordance with the I/O operation (4252), a metadata update (4254) describing a mapping of segments of content to virtual addresses within a storage object (4256), wherein the metadata update (4254) describes a mapping of segments of content to virtual addresses within a storage object (4256). The example method illustrated in FIG. 42B is similar to the example method illustrated in FIG. 42A because it also includes determining that the first storage system (4200A-4200N) includes a dataset (4258), and synchronizing (4206) the metadata at the second storage system (4200B) of storage systems (4200A-4200N) by sending metadata updates (4254) to the second storage system (4200B) to update the metadata representation on the second storage system according to the metadata updates (4254).

[0580] However, the example method shown in FIG. 42B further includes applying (4288) the I/O operation (4252) to the dataset (4258) on the first storage system (4200A); and, in response to successfully applying the I/O operation (4252) on the first storage system (4200A), updating (4290) a version of the metadata corresponding to the dataset (4258) at the first storage system (4200A) and in accordance with the metadata update (4254) before applying the I/O operation (4252); and determining (4292) predicate metadata describing the ordering of the I/O operation (4252) relative to one or more other I/O operations.

[0581] Applying (4288) the I/O operation (4252) to the dataset (4258) on the first storage system (4200A) may be performed by a controller of the storage system (4200A), as described above with respect to FIG. 1, and describes the controller performing the write operation using one or more of the memory components of the storage system (4200A), such as NVRAM, and persistent storage, such as flash memory, or any type of solid-state non-volatile memory.

[0582] Updating (4290) the version of the metadata corresponding to the dataset (4258) at the first storage system (4200A) and in accordance with the metadata update (4254) before applying the I/O operation (4252) may be performed by identifying a portion of the metadata representation for the corresponding storage object (4258), i.e., the source volume, and storing the dataset (4258), which portion may in some cases be the entire source volume. Further, the portion of the metadata representation for the storage object (4256) may be identified by traversing the structured collection of metadata objects described above using the memory address data for the I/O operation (4252) to find the node corresponding to the data object at the memory address for the I/O operation (4252). Additionally, a metadata object root node for the metadata representation may be created that references one or more nodes in the metadata representation for the storage object (4256), and the metadata object root node may specify a portion of one or more nodes in the metadata representation for the entire storage object (4256), or may specify other indications to reference only the portion of the metadata representation for the entire storage object (4256) that corresponds to the I/O operation (4252). In this manner, the metadata representation of the dataset reflects the successful application of the I/O operation (4252).

[0583] Determining predicate metadata (4292) describing the ordering of an I/O operation (4252) relative to one or more other I/O operations may be performed by tracking each received I/O operation and determining whether any dependencies exist between the I/O operations; after such I/O operations are identified, any ordering consistency issues should be resolved, as they should be during normal execution using techniques such as reader-defined ordering or predicates or through interlock exceptions. Interlock exceptions are described in application Ser. No. 15/696,418, which is incorporated herein by reference in its entirety. With respect to predicates, the description of the relationship between an operation and a common metadata update may be described as a set of interdependencies between separate modification operations—these interdependencies may be described as a set of precursors on which an operation depends in some way, and the set of precursors may be viewed as predicates that must be true for the operation to complete. Furthermore, predicates may not need to be protected if they are used to impose limits on concurrency between leaders and followers, as the persisted information could imply a variety of plausible conclusions if those predicates change the order in which the storage system persists the information.

[0584] For additional explanation, FIG. 43 sets forth a flowchart illustrating an example method for determining active membership among storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (4300A-4300N) illustrated in FIG. 43 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3C, or any combination thereof. In practice, the storage systems (4300A-4300N) illustrated in FIG. 43 may include the same components as the storage systems described above, fewer components, or additional components.

[0585] In the following example, determining how to recover from an error among storage systems that synchronously replicate a dataset may include determining whether to engage a quorum protocol. As described below, determining active membership within a storage system cluster or pod may overcome errors, such as a communication failure or storage device failure, by determining a subset of storage systems that continue to service I/O operations directed to the dataset, while another subset of storage systems does not continue to service I/O operations directed to the dataset. In this manner, either through intermediation or quorum policies, one or more storage systems may be determined to control the history of I/O operations that modify a synchronously replicated dataset as storage systems are added or removed from the in-sync list of storage systems.

[0586] The process of engaging an intermediary service in response to an error, such as a communication failure between storage systems—storage systems may be configured to store handles indicating contact addresses over a wide area network and cryptographically secure tokens that can be used to manage a pool of keys for the intermediary—is described in more detail in Application Serial No. 14/703,559, which is incorporated herein by reference in its entirety. Also described in Application Serial No. 15/703,559 is the use of various quorum protocols to determine which storage system in a collection of storage systems replicating a dataset should continue to service I/O requests directed to the dataset.

[0587] While this application Ser. No. 14/703,559 describes implementations for arbitration and quorum protocols, the focus of this disclosure is an analysis for determining which techniques to implement to determine active membership in a storage system cluster. For example, in some situations, one or more storage systems may engage in arbitration through one or more other storage systems and win arbitration—if the other one or more storage systems win arbitration, the storage system's performance for servicing I/O operations will be better according to one or more data storage metrics or performance criteria. In other words, it is not necessarily advantageous for one or more systems to win arbitration. Described below are techniques for avoiding situations where one or more storage systems that should not win arbitration win arbitration, including allocating a greater number of votes to storage systems that exhibit higher performance, or are more closely connected to a particular host system, or have other relatively better metrics than the number of votes allocated to storage systems with poorer metrics.

[0588] In some examples, to determine active membership in a storage system cluster, a default technique for determining a set of one or more storage systems to continue servicing I/O requests directed to a dataset may be implemented by using a quorum protocol—unless the storage system (4300A) can determine or prove that use of the quorum protocol would not allow for the establishment of a quorum for determining a set of one or more storage systems to continue servicing I/O requests. In other words, in response to an error, such as a communication failure between storage devices in the storage system (4300A), the controller of the storage system (4300A) may determine whether a quorum can be established—if a quorum can be established under the particular quorum protocol, then the quorum protocol is used to determine active or synchronizing membership in the storage system cluster; otherwise, if a quorum cannot be established under the particular quorum protocol, then the storage system (4300A) may engage in arbitration to determine active or synchronizing membership in the storage system cluster. In a simple case, if there are two storage systems that are members of the in-sync list, then quorum analysis is not performed in response to a communication failure because no other single storage system can form a quorum. In some examples, a given storage system or set of storage systems that communicate with each other may rely on a quorum policy to detach one or more storage systems that are not communicatively coupled unless the storage system or set of storage systems that communicate with each other can eliminate the possibility that another set of storage systems could form a quorum based on a comparison of votes for each set of storage systems. In other words, if one or more storage systems that are in communication with each other are able to form a quorum, then the one or more storage systems that are not communicating may be detached using the quorum policy without resorting to an intermediary.

[0589] In some examples, the controller for storage system 4300A may determine whether a quorum can be established by determining the set of storage systems that are in communication with each other - which also serves to determine the set of storage systems with which it is not in communication based on storage system 4300A referencing the in-sync list of storage systems 4300B-4300N that are included in the complete set of storage systems synchronously replicating the dataset. Additionally, storage system 4300A may refer to a stored indication of the number of votes each storage system in the set of all storage systems or the last known set of in-sync storage systems 4300A-4300N has with respect to a given quorum protocol. In this manner, given a determined first set of storage systems that are in communication with each other, along with respective votes corresponding to each of the first set of storage systems, and a determined second set of storage systems that are not in communication with the storage system (4300A) performing the analysis, along with respective votes corresponding to each of the second set of storage systems, the storage system (4300A) may determine whether a storage system in the first set of storage systems has sufficient votes to establish a quorum and whether a storage system in the second set of storage systems is likely to have sufficient votes to establish a quorum.

[0590] In other words, a storage system (4300A) detecting and responding to a communication failure may determine (1) whether storage systems that are not communicating can possibly form a quorum, (2) whether storage systems that remain in communication can possibly form a quorum, and/or (3) whether any storage systems that are not communicating can possibly determine whether storage systems that are communicating can possibly form a quorum—depending at least on which sets of storage systems remain in communication, which sets of storage systems are not communicating, and the respective votes corresponding to each storage system in each set of storage systems.

[0591] Furthermore, by ensuring that storage system (4300A) ensures that another storage system or another set of storage systems will not be able to form a quorum, storage system (4300A) ensures that if it wins arbitration along with one or more other storage systems that communicate with storage system (4300A), the other storage systems or other set of storage systems will be able to create a synchronously replicated version of the data set so that there will be no inconsistencies in the data set when one or more of the storage systems resynchronize with the storage system that won arbitration.

[0592] In one example, there may be an even number of votes associated with storage systems belonging to the in-sync list as members of a pod prior to some system failure. In this example, if a first set of in-sync pod member storage systems are in communication with each other and the first set of storage systems represent exactly half of the votes to establish a quorum, other sets of storage systems that may have been in communication with each other—but not with the first set of storage systems—would not be able to constitute more than half of the votes required to establish a quorum. In this example, a storage system included in the first set of storage systems may determine that neither the first set of storage systems nor the other sets of storage systems could possibly form a quorum, and in response, the storage system may initiate mediation.

[0593] As another example, the determination of whether a quorum can be established by communicating or non-communicating systems may be iteratively determined in response to multiple respective failure events—a different response may be performed in response to each failure event, such as by responding by engaging in mediation or quorum voting. For example, prior to a failure, such as a communications failure, a collection of storage systems may be in-syncing members of a pod. In this example, a particular storage system with one vote to establish a quorum may lose communication with other members of the pod, and in response, the other members of the pod, having enough votes to establish a quorum, remove the particular storage system through voting. After completion of this quorum voting phase to remove the particular storage system, the in-syncing members of the pod are the in-syncing systems of the pod excluding the particular storage system—resulting in an in-syncing member list of storage systems with a total of four votes to establish a quorum. To continue with this example, as described above, if a failure causes a storage system to belong to a collection of storage systems with a total of two votes, then the storage system may determine that quorum is not possible and may initiate mediation.

[0594] In some examples, storage systems that are in-sync members of a pod may be assigned different respective numbers of votes, including zero votes. For example, for a collection of storage systems that are in-sync members of a pod, different vote distributions include (a) all storage systems with a single vote, (b) some storage systems with multiple votes and some with a single vote, (c) some storage systems with multiple votes, some with multiple votes, and some with zero votes, or (d) each storage system with a different number of tables. In other words, generally, any given storage system that is an in-sync member of a pod may be assigned any number of tables, including zero votes. An example of a storage system with zero votes may occur during the migration of a data set from a source storage system that is a member of a pod to a target storage system that is not yet within the pod—before the migration is complete, the source storage system controls one or more of its votes and the target storage system controls none of the votes; after the migration is complete, the target storage system may be given control of one or more votes, and the source storage system may be updated to control or have no votes.

[0595] In some examples, an in-sync list may be established among a collection of storage systems as additional storage systems are added to or detached from a pod, where each storage system may maintain metadata indicating its membership in the pod, and the in-sync list further indicates the status of each member of the pod. When a storage system is added to a pod, the added storage system may be given an intermediation handle by the existing storage systems in the pod to contact the intermediation service. Furthermore, as changes are made to the pod, the in-sync list is updated to reflect the current membership of storage systems in the current pod. Additional discussion regarding pod definition and management can be found in Application Serial Nos. 62/470,172 and 62/518,071, which are incorporated herein by reference in their entirety. Furthermore, as a pod is expanded or extended to include multiple storage systems, the storage systems within the pod may be configured to request intermediation from a specific intermediation service—in this way, each storage system in the pod requests intermediation from the same intermediation service if intermediation is determined to be a response to a given error. Configuration of the storage system for accessing the intermediary service is further described herein.

[0596] As shown in FIG. 43, multiple storage systems (4300A-4300N) synchronously replicating a dataset (4352) may communicate with each other storage system via one or more networks (not shown) and with an intermediary service (4301) that may determine which storage system will continue to service the dataset in the event of a communication failure between the storage systems, if a storage system goes offline, or due to some other triggering event. In general, any number of storage systems may be part of the in-sync list that are synchronously replicating dataset (4352).

[0597] The example method shown in FIG. 43 includes detecting (4302) a communication failure involving loss of communication with one or more of the storage systems (4300B-4300N) by a particular storage system (4300A) of the storage systems (4300A-4300N), where the particular storage system (4300A) is configured to request intermediation from an intermediation target, such as an intermediation service (4301). Detecting (4302) the communication failure may be implemented using several techniques, including the controller of the particular storage system (4300A) not receiving communication from another storage system (4300B-4300N) over a communication link (4354) or channel for some period of time. In another example, detecting a communication failure (4302) may be performed by a controller of a particular storage system (4300A) according to a clock exchange protocol to determine that a communication channel is not operating correctly, clocks being described in more detail in other sections contained herein. Other standard techniques for detecting a communication failure (4302) may also be implemented.

[0598] The example method of FIG. 43 also includes determining (4304) that at least one of the one or more storage systems (4300B-4300N) is configured to request mediation from an intermediation target, such as an intermediation service (4301), in response to a communication failure. Determining (4304) that at least one of the one or more storage systems (4300B-4300N) is configured to request mediation from an intermediation target, such as an intermediation service (4301), in response to a communication failure may be performed by the storage system (4300A) undergoing the analysis described above - based on a comparison of votes corresponding to one or more systems in communication with the storage system (4300A) and votes corresponding to one or more storage systems not in communication with the storage system (4300A), the storage system (4300A) can determine whether one or more storage systems not in communication with the storage system (4300A) can form a quorum and whether those one or more storage systems not in communication with the storage system (4300A) are likely to be engaged in mediation with the intermediation target.

43 also includes determining (4306) whether to request intermediation from the intermediary target in response to determining (4304) that one or more of the storage systems (4300B-4300N) are configured to request intermediation from the intermediary target. Determining (4306) whether to request intermediation from the intermediary target may be performed based on the analysis and the determination (4304) that at least one of the one or more storage systems with which communication has been lost may be requesting intermediation from the intermediary target—if the storage system (4304A) determines that at least one of the one or more storage systems may be requesting intermediation from the metadata target, then the storage system (4300A) may also request intermediation from the intermediary target. Otherwise, storage system (4300A) and any storage systems in communication with storage system (4300A) may engage in a quorum policy to detach at least one of the one or more storage systems, or storage system (4300A) and any storage systems in communication with storage system (4300A) may determine that another set of storage systems may have a quorum and cease operating with copies of the synchronously replicated data set (those storage systems effectively go offline until they can establish communication with multiple storage systems, which may help determine the state of the in-sync and out-of-sync lists of storage systems).

[0600] For further explanation, Figure 44 illustrates a flowchart showing an example method for determining active membership between storage systems that synchronously replicate a database in accordance with some embodiments of the present disclosure. The example method illustrated in FIG. 44 is similar to the example method illustrated in FIG. 4 because it also includes detecting (4302) a communication failure with one or more of the storage systems (4300B-4300N) by a particular storage system (4300A) of the storage systems (4300A-4300N), where the particular storage system (4300A) is configured to request intermediation from an intermediation target; determining (4304) that at least one of the one or more storage systems (4300B-4300N) is configured to request intermediation from an intermediation target, such as an intermediation service (4301), in response to the communication failure; and determining (4306) whether one or more of the storage systems (4300B-4300N) is configured to request intermediation from the intermediation target in response to determining (4304) that the one or more of the storage systems (4300B-4300N) is configured to request intermediation from the intermediation target.

[0601] However, the example shown in FIG. 44 further includes determining, by a first set of one or more storage systems, that there is consistent communication among the storage systems of the first set of one or more storage systems (4402), where each storage system of the storage systems corresponds to zero or more votes in a quorum protocol that determines whether the first set of one or more storage systems may detach a second set of one or more storage systems; determining, by the first set of one or more storage systems, a lack of communication with a storage system among the second set of one or more storage systems (4404), where the first set of one or more storage systems cannot form a quorum; and determining, by the first set of one or more storage systems, that the second set of one or more storage systems also cannot form a quorum (4406).

[0602] Determining by the first set of one or more storage systems that there is consistent communication among the storage systems of the first set of one or more storage systems (4402) may be performed by each storage system of the first set of one or more storage systems exchanging status messages with every other system of the first set of one or more storage systems. In another example, a given storage system may determine that all storage systems that have consistent communication with each other are those storage systems in the in-sync list, with the exception of one or more storage systems for which a communication failure has been detected.

[0603] Determining (4404) a lack of communication with a storage system between a second set of one or more storage systems by a first set of one or more storage systems may be similarly performed by detecting a communication failure (4302), as described above with respect to FIG. 43.

[0604] Determining (4406) by a first set of one or more storage systems that a second set of one or more storage systems cannot also form a quorum may be performed as described above with respect to whether one set of storage systems can establish that another set of storage systems can form a quorum.

[0605] For further explanation, FIG. 45 sets forth a flowchart illustrating an example method for determining active membership among storage systems synchronously replicating a dataset in accordance with some embodiments of the present disclosure. The example method illustrated in FIG. 45 includes detecting (4302) a communication failure with one or more of the storage systems (4300B-4300N) by a particular storage system (4300A) among the storage systems (4300A-4300N), where the particular storage system (4300A) is configured to request intermediation from an intermediation target; determining (4304) that at least one of the one or more storage systems (4300B-4300N) is configured to request intermediation from an intermediation target, such as an intermediation service (4301), in response to the communication failure; and determining (4305) by a first set of one or more storage systems that a consistency check is performed among the storage systems of the first set of one or more storage systems. The example method illustrated in FIG. 45 is similar to the example method illustrated in FIG. 44 in that it also includes determining (4402) that there is communication with a storage system among a second set of one or more storage systems, by the first set of one or more storage systems, determining (4404) that there is a lack of communication with a storage system among a second set of one or more storage systems, by the first set of one or more storage systems, determining (4406) that the second set of one or more storage systems will not have a sufficient number or votes to form a quorum, and determining whether to request intermediation from the intermediation target in response to determining (4304) that one or more of the storage systems (4300B-4300N) are configured to request intermediation from the intermediation target.

[0606] However, the example method shown in FIG. 45 further specifies that determining (4404) by the first set of one or more storage systems a lack of communication with a storage system among the second set of one or more storage systems further includes determining (4502) that the first set of one or more storage systems cannot form a quorum according to which the first set of one or more storage systems includes exactly half of the votes for storage systems that are synchronously replicating the data set.

[0607] Determining (4502) that a quorum cannot be formed according to a first set of one or more storage systems including exactly half of the votes for storage systems synchronously replicating the first data set may be performed as described above with respect to whether one set of storage systems can confirm that another set of storage systems can form a quorum; if the first set of one or more storage systems controls exactly half or more than half of the total quorum votes, then the remaining storage systems will not be able to establish a quorum due to the lack of votes.

[0608] For further explanation, FIG. 46 sets forth a flowchart illustrating an example method for determining active membership among storage systems that synchronously replicate a dataset in accordance with some embodiments of the present disclosure. The example method illustrated in FIG. 46 includes detecting (4302) a communication failure with one or more of the storage systems (4300B-4300N) by a particular storage system (4300A) among the storage systems (4300A-4300N), where the particular storage system (4300A) is configured to request intermediation from an intermediation target; determining (4304) that at least one of the one or more storage systems (4300B-4300N) is configured to request intermediation from an intermediation target, such as an intermediation service (4301), in response to the communication failure; and determining (4305) by a first set of one or more storage systems that synchronously replicate a data set among the storage systems of the first set of one or more storage systems. The example method illustrated in FIG. 46 is similar to the example method illustrated in FIG. 44 in that it also includes determining, by a first set of one or more storage systems, that there is consistent communication between the storage systems (4300B-4300N) (4402); determining, by a first set of one or more storage systems, a lack of communication with a storage system between a second set of one or more storage systems (4404); determining, by the first set of one or more storage systems, that the second set of one or more storage systems will also be unable to form a quorum (4406); and determining whether to request intermediation from the intermediation target in response to determining (4304) that one or more of the storage systems (4300B-4300N) are configured to request intermediation from the intermediation target (4306).

[0609] However, the example method shown in FIG. 46 further specifies that determining (4406) by the first set of one or more storage systems that the second set of one or more storage systems also cannot form a quorum further includes determining (4602) that the second set of one or more storage systems would not be able to form a quorum according to the second set of one or more storage systems including half of the votes for the storage systems that synchronously replicate the dataset.

[0610] Determining (4602) that the second set of one or more storage systems will not be able to form a quorum according to the second set of one or more storage systems including half of the votes for the storage systems that synchronously replicate the data set may be performed as described above with respect to whether one set of storage systems can establish that another set of storage systems can form a quorum.

[0611] For further explanation, FIG. 47 sets forth a flowchart illustrating an example method for determining active membership among storage systems that synchronously replicate datasets in accordance with some embodiments of the present disclosure. While not shown in great detail, the storage systems (4300A-4300N) illustrated in FIG. 47 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3C, or any combination thereof. In practice, the storage systems (4300A-4300N) illustrated in FIG. 47 may include the same components as the storage systems described above, fewer components, or additional components.

[0612] The example method shown in FIG. 47 includes detecting (4702), by a first storage system of the storage systems, a communication failure with a second storage system of the storage systems. Detecting (4702), by a first storage system of the storage systems, a communication failure with a second storage system of the storage systems may be implemented similarly to detecting (4302), by a particular storage system (4300A) of the storage systems (4300A-4300N) a communication failure with one or more of the storage systems (4300B-4300N), as described above with respect to FIG. 43.

47 also includes generating (4704) an arbitration analysis (4752) indicating that winning arbitration will enable the storage systems in communication with the first storage system to continue servicing I/O requests directed to the dataset (4352) such that performance criteria will be better met compared to the second storage system continuing to service the I/O requests. Generating (4704) the arbitration analysis (4752) may be performed by the storage system (4300A) comparing one or more performance characteristics of the first storage system and one or more storage systems in communication with the first storage system to one or more corresponding performance characteristics of one or more storage systems not in communication with the first storage system to determine which set of storage systems will more effectively, efficiently, or reliably process the I/O requests. For example, a storage system may track multiple metrics that indicate performance with respect to processing I/O requests, and the metrics may be affected based on processor speed, network latency, read latency, and write latency, among other factors.

[0614] In another example, the mediation analysis may reflect a determination of whether a quorum can be established at a non-communicating storage system, where the mediation analysis indicates that storage system (4300A) engages in a mediation or quorum protocol, as described above with respect to FIG. 43.

[0615] The example method shown in FIG. 47 also includes requesting (4706) mediation from the mediation target (4301) according to the mediation analysis. Requesting (4706) mediation from the mediation target (4301) according to the mediation analysis may be performed as described above with respect to requesting mediation and as further described with respect to Application Serial No. 14/703,559.

[0616] For further explanation, FIG. 48 sets forth a flowchart illustrating an example method for synchronizing metadata between storage systems that synchronously replicate datasets, according to some embodiments of the present disclosure. While not shown in great detail, the storage system (4800A) shown in FIG. 48 may be similar to the storage systems described above with respect to FIGS. 1A-1D, 2A-2G, 3A-3C, or any combination thereof. In practice, the storage system (4800A) shown in FIG. 48 may include the same components as the storage systems described above, fewer components, or additional components.

[0617] In these examples, pod membership may be defined using a list of storage systems, a subset of which may be inferred to be synchronized, or in sync, for the pod. In some cases, the subset of the list may include every one of the storage systems for the pod, and the list may be considered metadata that is common to all storage systems and is consistently maintained across the pod through the use of one or more consistency protocols that are applied in response to changes in pod membership. A 'pod,' as the term is used here and throughout the rest of this application, may be embodied as a management entity representing a dataset, a collection of managed objects and management operations, a collection of access operations to modify or read the dataset, and multiple storage systems. Such management operations may modify or query managed objects equally well through any of the storage systems, and access operations to read or modify the dataset equally well through any of the storage systems. Each storage system may store a separate copy of the dataset as an appropriate subset of the dataset stored and advertised for use by the storage system, and operations that modify a management object or dataset performed and completed through any one storage system are reflected in subsequent management objects to query the pod or subsequent access operations to read the dataset. Additional details regarding "pods" may be found in previously filed provisional patent application Ser. No. 62/518,071, which is incorporated herein by reference.

[0618] A storage system may be considered in sync for a pod if the pod is at least in recovery with identical idle content for the last read copy of the dataset associated with the pod. Idle content is the content after any ongoing modifications have completed without processing new modifications. In some cases, this may be referred to as "crash-recoverable" consistency. Pod recovery may be considered the process of reconciling differences in applying concurrent updates to synchronized storage systems within the pod. Recovery may resolve any inconsistencies between storage systems in the completion of concurrent modifications that were requested for various members of the pod but were never signaled to any requester as successfully completed.

[0619] Given the use of a list of storage systems for a pod, a storage system that is listed as a pod member but not listed as in sync for the pod may be considered detached from the pod. Conversely, using the list of storage systems for a pod, a storage system that is listed as a pod member and is also listed as in sync and currently available to actively provide data for the pod may be considered online for the pod. Additionally, each storage system in a pod may have its own copy of the membership list, including which storage systems it last knew were in sync and which storage systems it last knew contained a complete set of pod members.

[0620] In this example, to be online for a pod, the membership list for a given storage system indicates that the given storage system is in sync for the pod—the given storage system can communicate with all other storage systems in the membership list that are indicated as in sync. If the storage system cannot confirm that it is in sync and in communication with all other storage systems in the membership that are indicated as in sync, then the storage system stops processing new incoming I/O commands or requests directed to the pod until the storage system can confirm that it is in sync and in communication with all other storage systems in the membership that are indicated as in sync. In some examples, if the storage system cannot confirm that it is in sync and in communication with all other storage systems in the membership that are indicated as in sync, then instead of stopping processing new I/O commands or requests, the storage system completes the I/O commands or requests with an error or exception. The I/O commands or requests may be SCSI requests, among other types of requests that use different network protocols. As an example, a first storage system may determine whether a second storage system in its membership list should be detached based on one or more criteria, and the result of the first storage system detaching the second storage system is that the first storage system continues to receive and process I/O commands because it is currently synchronized with at least all of the storage systems that remain on the membership list after the first storage system removed the second storage system from the membership list. However, to avoid a "split-brain" situation that could lead to incompatible datasets, dataset corruption, or application corruption, among other risks, the second storage system must be prevented from detaching the first storage system so that the second storage system—in addition to the first storage system—continues to receive and process I/O commands directed to datasets for the pod. In other words, if two different storage systems in a pod think they have successfully detached from each other, then a split-brain situation may result.

[0621] Situations for determining how to proceed when not communicating with another storage system in the membership list that is indicated as in sync may occur while the storage system is operating normally and notices one or more lost communications, may occur while the storage system is recovering from a previous failure, may occur when the storage system is switching operation from one set of storage system controllers to another set of storage system controllers for any reason, may occur during storage system startup or when a network interface is connected or enabled, or may occur during or after any combination of these and other types of events. In other words, whenever a storage system associated with a pod cannot communicate with all known, non-detached members of the membership list, the storage system may either wait, or go offline and possibly continue to wait, for example, some predetermined amount of time, until communication can be established, or the storage system may determine that it is safe to detach the non-communicating storage system without risking incurring a split-brain situation. Furthermore, if the safe detach occurs quickly enough, the storage system may remain continuously online for the pod with only a short delay and few or no failure requests, or some requests may result in a "busy" or "try again" that can be recovered through lower-level requester-side operation processing, with no adverse impact to applications or other high-level operations.

[0622] In some circumstances, a given storage system in a pod may determine that it is outdated or differently configured with respect to other storage systems in the pod after it is initially added to the pod with which it is in sync; the given storage system may determine that the existing storage systems in the pod are configured with newer or different software, firmware, or hardware combinations than the given storage system. As another example, a given storage system may determine that it is outdated or differently configured in response to the given storage system reconnecting to another storage system, and the other storage system marks the given storage system as detached - in this case, the given storage system may wait until it connects to some other set of storage systems that are in sync for the pod.

[0623] In these examples, it may be determined whether the manner in which storage systems are added or removed from a pod or from a synchronizing membership list can avoid transient inconsistencies. For example, transient inconsistencies may arise because each storage system has its own copy of the membership list, and because two or more independent storage systems in a pod may update their respective membership lists at different times—at least at times other than exactly simultaneously—and because a local copy of the membership list, possibly inconsistent with other membership lists, may account for all membership information that a given storage system may make available. As an example, if a first storage system is in sync for a pod and a second storage system is added, and the second storage system is then updated to list both the first and second storage systems as in sync in its respective membership list—before the first storage system lists the first and second storage systems as in sync in its respective membership list—if a failure occurs, causing both the first and second storage systems to restart, the second storage system may start up and wait to connect to the first storage system. Meanwhile, the first storage system may not be aware that it must or could wait for the second storage system. Continuing with this example, if the second storage system then responds to its inability to connect with the first storage system by going through a process to detach the first storage system, then the second storage system may succeed in completing a process that the first storage system is unaware of, resulting in a split-brain situation.

[0624] As an example technique for preventing the situation described in the example above, storage systems within a pod may follow a policy whereby individual storage systems will not disagree regarding whether they may choose to undergo the detachment process if they are not communicating. An example technique for ensuring that individual storage systems do not disagree is to ensure that when a new storage system is added to the in-sync membership list for a pod, the new storage system first remembers that the new storage system is a detached member. At this point, the existing in-sync storage system may locally store an indication that the new storage system is an in-sync pod member before the new storage system locally remembers that the new storage system is an in-sync pod member. Consequently, if there is a reboot or network failure or outage before the new storage system remembers the in-sync status for itself, then the original storage system—the storage system that is an in-sync member of the pod—may detach the new storage system due to the lack of communication before attempting to add the new storage system while the new storage system waits.

[0625] Continuing with this example, a reverse version of such a membership change may be required to remove a communicating storage system from a pod—first the storage system being removed or detached stores an indication locally that it is not in sync, and the storage system remaining in the pod subsequently stores an indication that the removed system is no longer in sync. At this point, both the storage system remaining in the pod and the storage system being removed remove the removed storage system from their respective membership lists. In this example, an intermediate, persisted detached state may not be required, depending on the implementation.

[0626] Furthermore, whether care is required in local copies of membership lists may depend on the use of model storage systems to monitor each other or to validate their membership. For example, if a consensus model is used for both, or if an external system—or an external distributed or clustered system—is used to store and validate pop membership, then discrepancies in locally stored membership lists may be inconsequential.

[0627] Some example models for resolving spontaneous membership changes include the use of a quorum, an external pod membership manager, or contention for known resources. These example models may be used in response to a communication failure, one or more storage systems in a pod failing, or an inability to communicate with a paired storage system in a pod, or a storage system starting up (or failing over to a secondary controller). Given these events that may trigger a change in pod membership, different membership models may use different mechanisms to define how a storage system in a pod decides to detach one or more paired storage systems in a safe manner and how to follow through when detaching one or more storage systems.

[0628] In some examples, there may be multiple membership lists used in reaching a consensus on membership changes. For example, for a given group of storage systems, each storage system may be on the in-sync list or the out-of-sync list, and each storage system stores a local copy of the in-sync and out-of-sync lists, respectively. In this example, the group of storage systems may be storage systems {A, B}, a pod may initially include storage system A, and the pod is stretched, or expanded, from storage system A to storage system B. This stretching of the pod is equivalent to expanding the storage system membership for the pod and may begin by ensuring that storage system A and storage system B are connected. Ensuring that storage system A and storage system B are connected may be a configuration step that precedes the stretch operation—however, mere connectivity between storage system A and storage system B does not stretch the pod; rather, connectivity between storage system A and storage system B allows the pod to stretch. In this example, storage system A may receive a command—e.g., from a management console for managing volumes, pods, and storage systems—indicating that a pod, i.e., a particular volume of a pod at storage system A, be stretched to storage system B. Given connectivity between storage system A and storage system B, the initial state may be described as storage system A storing an in-sync list denoting {A} and an out-of-sync list denoting {B}, where the epoch identifier is equal to n, the membership sequence is equal to m, and storage system B storing empty lists for both the in-sync and out-of-sync lists. In response to storage system A receiving the stretch command, storage system A may send a message to storage system B indicating the session identified by the pod identifier, epoch identifier n, and in response, storage system B communicates back to storage system A. Further, the configuration-level heartbeat between storage systems A and B may distribute the in-sync and out-of-sync lists for storage system A to storage system B, in response storage system B may determine that it is not an in-sync member and may initiate a re-sync operation with storage system A, which synchronizes the pod across both storage system A and storage system B. Further, in response to the re-sync, storage system A may write an updated in-sync list {A, B} to storage system B and then wait for storage system B to respond. At this point, storage system A is ready to begin communication with storage system B regarding the in-sync operation - however, storage system B will not engage in such communication until storage system B receives the updated in-sync list {A, B} as an in-sync member of the pod. For example, storage system A may begin communication by initiating a clock exchange operation with storage system B, but storage system B may not initiate the clock exchange operation until storage system B receives the pending in-sync list {A, B}. Clock switching is described in more detail in application reference numbers 62/470,172 and 62/518,071, which are incorporated herein in their entireties.

[0629] Continuing with this example, to unstretch, i.e., remove a storage system from membership in a pod, a member storage system may take the following steps. For example, if the pod membership is currently {A, B}, where both storage system A and storage system B have the same in-sync list of {A, B} and out-of-sync list of { }, a current epoch of n, and a current membership sequence of m—in this situation, storage system A may receive a request to unstretch the pod and remove storage system B. In response to the unstretch request, storage system A may send a message to storage system B indicating the committed membership list, indicating an in-sync list of {A, B} and an out-of-sync list of { }, and a pending membership list indicating an in-sync list of {A} and an out-of-sync list of { }, a current epoch of n, and a membership sequence of (m+1). In response to receiving the message from storage system A, storage system B applies the state information indicated in the message and responds to storage system A that the state change has been applied. In response to receiving an acknowledgment of the state change from storage system B, storage system A updates its local state information to indicate committed membership lists for the in-sync list of {A} and the out-of-sync list of {}, an in-sync list of {A} and an out-of-sync list of {}, and a pending membership list for epoch (n+1), and storage system B then stops communicating with storage system B. Storage system B detects the lost session, but because it has an in-sync list of {A}, it requests a re-establishment of the session from storage system A and receives a response indicating that storage system B is no longer a member of the pod.

[0630] In examples using quorum as a membership model, one technique for resolving detach operations is to use a majority—or quorum—model for membership. For example, given three storage systems, as long as two are communicating, the two communicating can agree to detach the third storage system that is not communicating. However, the third storage system cannot independently choose to detach one of the two communicating storage systems. In some cases, confusion can arise when storage system communication within a pod is inconsistent. In this example, for storage systems {A, B, C}, storage system B may be in communication with both storage system A and storage system C, while storage system A may be in communication with storage system B but not with storage system C. In this situation, both storage systems A and B may detach storage system C—or both storage systems B and C may detach storage system A—but more communication between pod members may be required to understand membership.

[0631] Continuing with this example, a quorum policy, or quorum protocol, may resolve this situation for adding or removing storage systems from a pod. For example, if a fourth storage system is added to a pod, then the majority of storage systems becomes three storage systems. Transitioning from a pod with three storage systems, where two are required for a majority, to a pod with four storage systems, where three are required for a majority, may require something similar to the model described above for carefully adding storage systems to the in-sync list. For example, a fourth storage system, say storage system D, starts in an attaching state but is not yet attached, and it will never generate more than a quorum of votes. Given that storage system D is in an attaching state, storage system A, storage system B, and storage system C may each be updated to recognize storage system D and updated with the new requirement for the three storage systems to reach a majority decision to detach any particular storage system from the pod. Additionally, removing a given storage system from a pod may similarly transition to a state in which the given storage system is detached before updating other storage systems in the pod. In some cases, a problem with the quorum model is that a common configuration is a pod with exactly two storage systems; in such cases, one solution is to add a storage system into the network that only participates in quorum votes for the pod but does not otherwise store the dataset for the pod. In this case, such a voting-only member would generally not trigger a series of quorum votes, but would only participate in votes triggered by storage systems in the pod configured as in-syncing storage systems.

[0632] In an example using an external pod membership method as the membership model, one technique involves using an external system outside the storage system itself to handle pod membership and manage membership transitions. For example, to become a member of a pod, a future storage system is configured to contact the pod membership system to request membership for the pod and verify that the future storage system is in sync for the pod. In this model, any storage system that is online, i.e., in sync, for a pod should remain in communication with the pod membership system and should wait, i.e., go offline, if communication with the pod membership system is lost. In this example, the pod membership system may be implemented as a highly available cluster using a variety of cluster tools, such as, for example, Oracle™ RAC, Linux HA, VERITAS™ Cluster Server, IBM™ HACMP, or others. In other examples, the pod membership system may be implemented using a distributed configuration tool, such as Etcd™ or Zookeeper™, or a trusted distributed database, such as DynamoDM™ by Amazon. Additionally, in other examples, pod membership may be determined using a distributed consensus algorithm, such as RAFT or PAXOS; implementations based on concepts from RAFT may include a RAFT-based internal algorithm for membership, or a RAFT-inspired algorithm for log-style update consistency that can be used as part of an overall solution for determining valid and up-to-date membership and for determining the current value of up-to-date membership information.

[0633] In an example using a race for resources, known as a membership model, or racing protocol, a cluster manager for a pod may resolve membership changes by requesting access to some resource that can be locked in some way to the exclusion of others, or by requesting access to a majority of several such resources. For example, one technique is to use a resource reservation, such as a SCSI reservation or SCSI persistent reservation, to obtain a lock on one or more networked SCSI devices. In this example, if a majority of the configured set of these networked devices can be locked by a storage system, then that storage system may detach other storage systems. Otherwise, the storage system would not be able to detach other storage systems. Furthermore, to stay online, i.e., synchronized, a storage system may need to frequently reassert or test its locks on resources, or be in communication with some other storage systems that are asserting, reasserting, or testing its locks on resources. Furthermore, networked computing resources that can be asserted and tested in various ways may be used similarly.

[0634] Continuing with this example, to ensure that a prolonged outage of all storage system members of a pod can be handled appropriately, while allowing one storage system to resume as a member pod and other storage system members to detach, network resources such as those described above must have persistent properties that may be used to verify that some other storage system has not previously detached the resuming storage system pod member. However, if the service only provides resource reservations without the ability to persistently store status information or other metadata, then the resource summary service may be used to gain access to some externally stored data, such as a third-party database or cloud storage, that may be queried and written after a particular storage system gains access—the data that is written may record information that a detached storage system may query to determine that it has been detached.

[0635] In some examples, the tracing protocol may be implemented using an intermediary service, a service that determines whether one storage system has the authority to detach another storage system from a pod. Example implementations of intermediary services are further described in Application Serial No. 15/703,559, which is incorporated herein by reference in its entirety.

[0636] In another example, a combination of mechanisms may be used that may be useful when a pod is spread across three or more storage systems. In one example, priority rules may be combined with metadata. In a top-of-rack example, a larger central storage system in a data center or campus may itself be synchronously replicated to a larger storage system in a second location. In that case, the top-of-rack storage system may never restart on its own and may give priority to either of the larger central storage systems in the two locations. The two larger storage systems in that case may be configured to mediate between each other, such that any smaller storage system that can connect to either of the two larger storage systems that remain online may continue to serve its pod, and any smaller storage system that cannot connect to either of the two larger storage systems (or can only connect to one that is offline for the pod) may cease serving the pod. Furthermore, the priority model may be combined with a quorum-based model. For example, three large storage systems in three locations might use a quorum model between each other, with smaller satellite or top-of-rack storage systems lacking any votes and only functioning if they can connect to one of the larger, synchronizing storage systems that is online.

[0637] In another example of combining mechanisms, brokering may be combined with a quorum model. For example, there may be three storage systems that vote among themselves to ensure that a third storage system, with which two storage systems normally do not communicate, can be safely detached. On the other hand, one storage system cannot absolutely detach two other storage systems on its own. However, after two storage systems successfully detach the third storage system, the configuration is now reduced to two storage systems that agree that they are in sync and agree on the fact that the third storage system is detached. In that case, the remaining two storage systems may agree to handle additional storage system or network failures using brokering (e.g., using a cloud service). This combination of brokering and quorum can be further extended. For example, in a pod spread across four storage systems, any three can detach the fourth storage system, but if two in-sync storage systems are communicating with each other but not with the other two storage systems, they will currently consider themselves both in sync and will then use brokering to safely detach the other two. Even in a five-storage-system pod configuration, if four storage systems vote to detach the fifth storage system, then the remaining four can use mediation if they split into two equal halves; once the pod is down to two storage systems, they can use mediation to resolve successive failures. Five to three might use quorum among them, then allow a drop to two, with the remaining two storage systems again using mediation if there are additional failures. This general multi-mode quorum mediation mechanism can handle an additional number of situations that neither quorum nor independent mediation between symmetric storage systems can handle. This combination can increase the number of cases where a faulty or occasionally unreachable mediator can be reliably used (or in the case of cloud mediators where customers may not fully trust them). Furthermore, this combination better handles the case of a three-storage-system pod, where the first storage system may successfully detach the second and third storage systems during a network failure that only affects the first storage system, as a result of the mediation alone. This combination may also better handle a series of failures affecting one storage system at a time, as illustrated in the three-to-two-to-one example. These combinations work because the in-sync and detach operations result in a specific state—in other words, moving from detached to in-sync is a process, and each stage in the quorum/mediator sequence ensures that all online/in-sync storage systems agree on the current, durable state for the pod, so the system is stateful. This differs from some other cluster models, where simply having a majority of cluster nodes in communication again is expected to be sufficient to resume operation. However, a priority model may still be added, where satellite or top-of-rack storage systems never participate in mediation or quorum and only serve pods if they can connect to online storage systems that participate in mediation or quorum.

48 includes determining (4802) that a membership event corresponds to a membership change for the set of storage systems (4800A-4800B) that synchronously replicate a data set (4858). Determining (4802) that a membership event corresponds to a membership change for the set of storage systems (4800A-4800B) that synchronously replicate a data set (4858) may be implemented using different techniques. As an example, storage system (4800A) may receive an I/O command indicating that pod (4854) is to be extended to include a new storage system (4800N) or that pod (4854) is to be de-extended to exclude an existing storage system (4800N). As another example, the storage system (4800A) may detect and determine that communication has been lost or has become unreliable or inefficient beyond a specified threshold with a particular storage system (4800N) in the collection of storage systems.

[0639] Receipt of an I/O command for a pod (4854) or an I/O operation (4852) directed to a dataset (4858) at a storage system (4800A) of a collection of storage systems (4800A-4800N) may be accomplished by using one or more communication protocols for moving packets or data over a network, such as a storage area network (158), the Internet, or any computer network over which a host computer (4851) can communicate with the storage system (4800A) in its entirety. In some cases, receiving I/O commands or I/O operations 4852 directed to a dataset 4858 for a pod 4854 may be performed using the communications interconnect 173 between the storage systems 4800A-4800N of the pod 4800A—or some other communications channel internal to the storage system 4800A—wherein the I/O commands or operations are received from applications or processes resident or running on the storage system computing resources. Additionally, applications, resident or remote, may use the storage systems 4800A-4800N in implementing file systems, data objects, databases, or other protocols that may provide functionality that relies on the storage systems 4800A-4800N being in sync and online—and any of these protocols or applications may be distributed implementations that operate on a synchronously replicated, symmetrically accessible underlying storage implementation. In this example, storage system (4800A) may receive an I/O command or I/O operation (4852) received at a network port, such as a SCSI port, where the I/O operation (4852) is a write command directed to a storage location that is part of a dataset (4858) that is synchronously replicated across storage systems (4800A-4800N) in the pod.

[0640] The example method shown in FIG. 48 also includes applying (4804) one or more membership protocols in accordance with the membership event to determine a new set of storage systems for synchronously replicating the data set (4858). Applying (4804) one or more membership protocols in accordance with the membership event to determine a new set of storage systems for synchronously replicating the data set (4858) may be implemented as described above using any one or more of a quorum protocol, an external pod membership manager protocol, or a racing protocol.

48 also includes, for one or more I/O operations (4852) directed to the dataset (4858), applying (4806) the one or more I/O operations (4852) to the dataset (4858) synchronously replicated by the new set of one or more storage systems. Applying (4806) the one or more I/O operations (4852) to the dataset (4858) synchronously replicated by the new set of storage systems may be implemented as described in Application Nos. 62/470,172 and 62/518,071, which are incorporated herein in their entirety and describe receiving and processing I/O operations such that any changes to the dataset are synchronously replicated across all synchronized storage system members of the pod.

[0642] The reader will understand that the above-described method may be implemented with any combination of the above-described storage systems. Furthermore, any of the above-described storage systems may be paired with storage provided by a cloud service provider, such as Amazon™ Web Services (“AWS”), Google™ Cloud Platform, Microsoft™ Azure, or others. Thus, in such an example, a member of a particular pod may include one of the above-described storage systems, as well as a storage system consisting of storage provided by the cloud service provider. Similarly, a member of a particular pod may consist solely of a logical representation of a storage system consisting of storage provided by the cloud service provider. For example, a first member of a pod may be a logical representation of a storage system consisting of storage in a first AWS availability zone, while a second member of the pod may be a logical representation of a storage system consisting of storage in a second AWS availability zone.

[0643] To facilitate the ability to synchronously replicate data sets (or other managed objects, such as virtual machines) to a storage system comprising storage provided by a cloud service provider and perform all other functions described herein, software modules implementing various storage system functions may execute on processing resources provided by the cloud service provider. Such software modules may execute on one or more virtual machines supported by the cloud service provider, such as block devices, Amazon™ Machine Image (“AMI”) instances, etc. Alternatively, such software modules may instead execute in a bare metal environment provided by the cloud service provider, such as an Amazon™ EC2 bare metal instance, where the hardware is directly accessible. In such an embodiment, an Amazon™ EC2 bare metal instance may be paired with high-density flash drives to effectively form a storage system. In either implementation, the software module would ideally be co-located on cloud resources with other traditional data sensor services, such as virtualization software and services offered by VMware™, such as vSAN™. The reader will understand that many other implementations are possible and within the scope of this disclosure.

[0644] The reader will understand that in situations where a dataset or other managed object within a pod is maintained in an on-premises storage system and the pod is expanded to include a storage system whose resources are provided by a cloud service provider, the dataset or other managed object may be transferred to the storage system whose resources are provided by the cloud service provider as encrypted data. Such data may be encrypted by the on-premises storage system, such that data stored in resources provided by the cloud service provider is encrypted, but the cloud service provider does not have the encryption key. In this manner, data stored in the cloud may be more secure because the cloud does not have access to the encryption key. Similarly, when data is initially written to an on-premises storage system, network encryption could be used, and the encrypted data could be transferred to the cloud such that the cloud does not subsequently have access to the encryption key.

[0645] Disaster recovery can be provided as a service through the use of a storage system consisting of storage provided by a cloud service provider. In such an example, datasets, workloads, other managed objects, etc. may reside on an on-premises storage system, and the resources may be synchronously replicated to a storage system provided by the cloud service provider. If a disaster occurs to the on-premises storage system, the storage system whose resources are provided by the cloud service provider may take over processing requests directed to the dataset, assist in migrating the dataset to another storage system, etc. Similarly, the storage system whose resources are provided by the cloud service provider may serve as an on-demand secondary storage system that may be used during periods of heavy use or otherwise as needed. The reader will understand that user interfaces or similar mechanisms for initiating many of the functions described herein can be designed, making enabling disaster recovery as a service as simple as performing a single mouse click.

[0646] High availability can also be provided as a service through the use of a storage system consisting of storage provided by a cloud service provider. In such an example, a data set, workload, or other managed object that may reside in an on-premises storage system may be synchronously replicated to a storage system whose resources are provided by the cloud service provider. In such an example, because of dedicated network connectivity to the cloud, such as AWS Direct Connect, sub-millisecond latency to AWS from various locations can be achieved. Thus, applications can be run in a stretched cluster mode without significant upfront expenditure, and high availability can be achieved without the need for multiple separately located on-premises storage systems to be purchased, maintained, etc. The reader will understand that user interfaces or similar mechanisms can be designed to initiate many of the functions described herein, thereby enabling applications to be extended into the cloud with a single mouse click.

[0647] System recovery can also be provided as a service through the use of a storage system consisting of storage provided by a cloud service provider. In such an example, point-in-time copies of databases, managed objects, and other entities that may reside in an on-premises storage system can be synchronously replicated to a storage system whose resources are provided by a cloud service provider. In such an example, if a need arises to restore a storage system to a particular point in time, the point-in-time copies of data sets and other managed objects contained on the storage system whose resources are provided by a cloud service provider can be used to restore the storage system.

[0648] Using a storage system consisting of storage provided by a cloud service provider, data stored in an on-premises storage system may be natively piped into the cloud for use by various cloud services. In such an example, data in its native format as it was stored in the on-premises storage system may be cloned and converted into a format usable for the various cloud services. For example, data in its native format as it was stored in the on-premises storage system may be cloned and converted into a format used by Amazon™ Redshift so that data analytics queries can be performed on the data. Similarly, data in its native format as it was stored in the on-premises storage system may be cloned and converted into a format used by Amazon™ DynamoDB, Amazon™ Aurora, or some other cloud database service. Because such conversion occurs outside of the on-premises storage system, resources in the on-premises storage system may be protected and preserved for use in servicing I/O operations. On the other hand, cloud resources that can be spanned as needed are used to perform data transformations, which can be particularly valuable in embodiments where an on-premises storage system acts as a primary servicer of I/O operations and a storage system comprised of resources provided by a cloud service provider acts more as a backup storage system. Indeed, because managed objects can be synchronized across storage systems, in embodiments where an on-premises storage system was initially responsible for performing the steps required in an extract-transform-write ("ETL") pipeline, components of such pipelines may be exported to the cloud and executed in the cloud environment. Use of such techniques may also provide analytics as a service, including using point-in-time copies (i.e., snapshots) of datasets as input to analytics services.

[0649] The reader will understand that applications can run on any of the storage systems described above, and that in some embodiments, such applications can run on the primary controller, the secondary controller, or on both controllers simultaneously. Examples of such applications may include applications that perform background batch database scans, applications that perform statistical analysis of runtime data, etc.

[0650] Example embodiments are described primarily in the context of a fully functional computer system. However, readers skilled in the art will recognize that the present disclosure may be embodied in a computer program product disposed on a computer-readable storage medium for use with any suitable data processing system. Such a computer-readable storage medium may be any storage medium for machine-readable information, including magnetic, optical, or other suitable media. Examples of such media include magnetic disks or diskettes in hard drives, compact discs for optical drives, magnetic tape, and others, as will occur to those skilled in the art. Those skilled in the art will readily recognize that any computer system with appropriate programming means is capable of executing the steps of the methods embodied in the computer program product. Those skilled in the art will also recognize that while some of the example embodiments described herein are directed to software installed and running on computer hardware, alternative embodiments implemented as firmware or as hardware are nevertheless well within the scope of the present disclosure.

[0651] Embodiments may include systems, methods, and/or computer program products. The computer program product may include a computer-readable storage medium (or media) having computer-readable program instructions thereon for causing a processor to implement aspects of the present disclosure.

[0652] A computer-readable storage medium may be a tangible device capable of retaining and storing instructions for use by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the above. A non-exhaustive list of more specific examples of computer-readable storage media includes the following: portable computer diskettes, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (PROM) (EROM or flash memory), static random access memory (SRAM), read-only compact disks (CD-ROMs), digital versatile disks (DVDs), memory sticks, floppy disks, mechanically encoded devices such as punch cards or ridge structures in grooves on which instructions are recorded, and any suitable combination of the above. Computer-readable storage media, as used herein, should not be construed as being primarily a primary signal, such as, for example, radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through a waveguide or other transmission medium (e.g., light pulses passing through a fiber optic cable), or electrical signals transmitted through wires.

[0653] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing device/processing device, or to an external computer or external storage device, for example, via the Internet, a local area network, a wide area network, and/or a wireless network. The network may include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and/or edge servers, etc. A network adapter card or network interface within each computing device/processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage on a computer-readable storage medium within the respective computing device/processing device.

[0654] Computer-readable program instructions for carrying out the operations of the present disclosure may be either source code or object code written in any combination of one or more programming languages, including assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state-setting data, or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Smalltalk, C++, and traditional procedural programming languages such as the "C" programming language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter situation, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be to an external computer (e.g., through the Internet using an Internet Service Provider). In some embodiments, electronic circuitry including, for example, a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA) may execute computer-readable program instructions by utilizing state information of the computer-readable program instructions to personalize the electronic circuitry to perform aspects of the present disclosure.

[0655] Aspects of the present disclosure are described herein with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems), and computer program products according to some embodiments of the present disclosure. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer-readable program instructions.

[0656] These computer-readable program instructions may be provided to a processor of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus to produce a machine, whereby the instructions executing on the processor of the computer or other programmable data processing apparatus create means for performing the functions/acts specified in one or more blocks of the flowcharts and/or block diagrams. These computer-readable program instructions may also be stored on a computer-readable storage medium that can instruct a computer, programmable data processing apparatus, and/or other device to function in a particular manner, whereby the computer-readable storage medium having instructions stored therein includes an article of manufacture containing instructions that implement aspects of the functions/acts specified in one or more blocks of the flowcharts and/or block diagrams.

[0657] Computer-readable program instructions may also be loaded into a computer, other programmable data processing device, or other device to cause the computer, other programmable device, or other device to execute a series of operational steps to produce a computer-implemented process, whereby the instructions executing on the computer, other programmable device, or other device perform the functions/acts specified in one or more blocks of the flowcharts and/or block diagrams.

[0658] The flowcharts and block diagrams depicted in the figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to aspects of the present disclosure. At this point, each block in the flowcharts or block diagrams may represent a module, segment, or portion of instructions, including one or more executable instructions for implementing the specified logical function(s). In some alternative implementations, the functions noted in the blocks may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may be executed in the reverse order, depending on the functionality involved. It should also be noted that each block in the block diagrams and flowchart diagrams, and combinations of blocks in the block diagrams and/or flowchart diagrams, may be implemented by a dedicated hardware-based system that performs the specified functions or acts, or a combination of dedicated hardware and computer instructions.

[0659] The reader will understand that the steps described herein may be implemented in a variety of ways, and that no particular order is required. It will also be understood from the foregoing description that modifications and variations may be made in various embodiments of the present disclosure without departing from its true spirit. The description herein is for illustrative purposes only and should not be construed in a limiting sense. The scope of the present disclosure is limited only by the language of the following claims.

Claims (20)

1. A method comprising:
In response to detecting, by one or more storage systems, a trigger event, requesting intermediation from an intermediation service external to the one or more storage systems, the intermediation service configured to determine which storage systems will continue to provide services to the data set after the trigger event;
processing data storage requests directed to the dataset by a first storage system that receives a positive intermediation result from the intermediation service, and storage systems that do not receive a positive intermediation result from the intermediation service not processing data storage requests directed to the dataset;
A method comprising: the trigger event is a communication failure between the first storage system and at least one other storage system among the one or more storage systems ;
The method of claim 1. the trigger event is the expiration of a synchronous replication lease;
The method of claim 1. detaching, by the first storage system, the at least one other storage system of the one or more storage systems from the one or more storage systems that synchronously replicate the data set in response to the indication of the positive mediation result from the mediation service;
The method of claim 2 further comprising: the request for mediation from each storage system is a request to establish a shared resource indicating an identifier for each requesting storage system;
The method of claim 1. the first storage system waits a first time period before requesting the intermediation;
the at least one other storage system of the one or more storage systems waits a second time period before requesting the intermediation.
The method of claim 4. the first storage system is designated as a preferred storage system based on the first time period being less than the second time period;
The method of claim 6. the intermediation service is a database service provided by a cloud service provider;
The method of claim 1. the intermediary service is implemented on a host computer that issues requests to modify data to the first storage system and at least one other storage system among the one or more storage systems ;
The method of claim 1. the intermediation service is implemented in a failure domain that is distinct from a failure domain of either the first storage system or at least one other storage system of the one or more storage systems ;
The method of claim 1. the intermediation service is one of a plurality of intermediation services;
The method of claim 1. The device is
A computer processor;
a computer memory operatively coupled to the computer processor, the computer memory having, when executed by the computer processor, a program for causing the apparatus to:
In response to detecting, by one or more storage systems, a trigger event, requesting intermediation from an intermediation service external to the one or more storage systems, the intermediation service configured to determine which storage systems will continue to service the data set after the trigger event;
processing data storage requests directed to the dataset by a first storage system of the one or more storage systems that receives a positive intermediation result from the intermediation service, and a storage system that does not receive a positive intermediation result from the intermediation service not processing data storage requests directed to the dataset;
said computer memory having computer program instructions disposed therein for executing
Equipped with. the trigger event is a communication failure between the first storage system and at least one other storage system among the one or more storage systems ;
13. The apparatus of claim 12. the trigger event is the expiration of a synchronous replication lease;
13. The apparatus of claim 12. When executed by the computer processor, the apparatus:
detaching, by the first storage system, at least one other of the one or more storage systems from the one or more storage systems that synchronously replicate the data set in response to the indication of the positive mediation result from the mediation service;
The apparatus of claim 12 , further comprising computer program instructions to execute: the request for mediation from each storage system is a request to establish a shared resource indicating an identifier for each requesting storage system;
13. The apparatus of claim 12. the first storage system waits a first time period before requesting the intermediation;
at least one other storage system of the one or more storage systems waits a second time period before requesting the intermediation;
13. The apparatus of claim 12. the first storage system is designated as a preferred storage system based on the first time period being less than the second time period;
18. The apparatus of claim 17. the intermediation service is a database service provided by a cloud service provider;
13. The apparatus of claim 12. 1. A method of mediating between a set of storage systems that synchronously replicate a data set, the set of storage systems including a first storage system and a second storage system;
detecting, by the first storage system, a communication failure between the first storage system and the second storage system;
requesting, by the first storage system, intermediation from an intermediation service in response to detecting the communication failure;
detaching, by the first storage system, the second storage system from the set of storage systems that synchronously replicate the data set in response to an indication of a positive mediation result from the mediation service;
A method comprising: